{"id":92253,"date":"2026-09-07T12:28:37","date_gmt":"2026-09-07T09:28:37","guid":{"rendered":"https:\/\/twodots.gr\/?p=92253"},"modified":"2026-09-07T12:28:37","modified_gmt":"2026-09-07T09:28:37","slug":"crate-elegchos-mobile-ai-agent-epitychia-asfaleia","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/crate-elegchos-mobile-ai-agent-epitychia-asfaleia\/","title":{"rendered":"CRATE: \u03c0\u03ce\u03c2 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5\u03bc\u03b5 \u03b1\u03bd \u03ad\u03bd\u03b1\u03c2 mobile AI agent \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03bc\u03b5 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u03a4\u03bf CRATE \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03b1\u03bd \u03ad\u03bd\u03b1\u03c2 mobile AI agent \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03bc\u03b5 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1, \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03bf\u03bd\u03c4\u03b1\u03c2 \u03ba\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1 \u03c3\u03b5 \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03bf evidence record \u03ba\u03b1\u03b9 \u03ba\u03c1\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03c2 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac \u03c4\u03b7\u03bd \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7 \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf.<\/strong> \u03a3\u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03b4\u03b7\u03bc\u03bf\u03c3\u03af\u03b5\u03c5\u03c3\u03b7 \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 F1 0,833 \u03c3\u03b5 116 AndroidWorld trajectories, \u03b5\u03bd\u03ce \u03c4\u03bf CRATE-S \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 F1 0,697 \u03c3\u03b5 204 MobileRisk trajectories. \u0393\u03b9\u03b1 production \u03c7\u03c1\u03ae\u03c3\u03b7, \u03cc\u03bc\u03c9\u03c2, \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 domain-specific \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 labels, privacy-safe logs \u03ba\u03b1\u03b9 \u03c3\u03b1\u03c6\u03ae approval gates.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#teliki-othoni-den-apodeiknyei-diadromi\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03bb\u03b7 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae<\/a><\/li>\n<li><a href=\"#proto-stadio-synepeies-ana-vima\">\u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf: \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b5\u03c2 \u03b1\u03bd\u03ac \u03b2\u03ae\u03bc\u03b1<\/a><\/li>\n<li><a href=\"#deytero-stadio-elegchos-oron-ergasias\">\u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf: \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03cc\u03bb\u03c9\u03bd \u03c4\u03c9\u03bd \u03cc\u03c1\u03c9\u03bd<\/a><\/li>\n<li><a href=\"#crate-s-asfaleia-kai-oloklirosi\">CRATE-S: \u03b7 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03b4\u03b5\u03bd \u03c4\u03b1\u03c5\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7<\/a><\/li>\n<li><a href=\"#benchmarks-pragmatikes-syskeves-etiketes\">\u03a4\u03c1\u03af\u03b1 benchmarks \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c5\u03c3\u03ba\u03b5\u03c5\u03ad\u03c2<\/a><\/li>\n<li><a href=\"#apotelesmata-task-completion\">\u03a4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c3\u03c4\u03bf task completion<\/a><\/li>\n<li><a href=\"#dokimes-operational-safety\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2 operational safety<\/a><\/li>\n<li><a href=\"#ablations-simasia-sympiesis\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03c4\u03c9\u03bd \u03b2\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd \u03ad\u03c7\u03b5\u03b9 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1<\/a><\/li>\n<li><a href=\"#business-control-layer\">\u0391\u03c0\u03cc \u03c4\u03bf benchmark \u03c3\u03b5 business control layer<\/a><\/li>\n<li><a href=\"#pilot-elegchou-mobile-ai-agent\">\u03a0\u03ce\u03c2 \u03c3\u03c4\u03ae\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 pilot \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2<\/a><\/li>\n<li><a href=\"#oria-paragogikis-yiothetisis\">\u03a4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae \u03c5\u03b9\u03bf\u03b8\u03ad\u03c4\u03b7\u03c3\u03b7<\/a><\/li>\n<li><a href=\"#axiologisi-agents-meros-proiontos\">\u0397 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2<\/a><\/li>\n<\/ul>\n<\/div>\n<p>\u03a0\u03ce\u03c2 \u03be\u03ad\u03c1\u03b5\u03b9\u03c2 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1\u03c2 AI agent \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03bc\u03b9\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c3\u03c4\u03bf \u03ba\u03b9\u03bd\u03b7\u03c4\u03cc \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 \u03b1\u03c0\u03bb\u03ce\u03c2 \u03c3\u03b5 \u03bc\u03b9\u03b1 \u03bf\u03b8\u03cc\u03bd\u03b7 \u03c0\u03bf\u03c5 \u03bc\u03bf\u03b9\u03ac\u03b6\u03b5\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ae; \u039a\u03b1\u03b9 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03c0\u03b9\u03bf \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc: \u03c0\u03ce\u03c2 \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9\u03c2 \u03cc\u03c4\u03b9 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03ba\u03ac\u03bb\u03c5\u03c8\u03b5 \u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b9\u03ba\u03ac \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, \u03b4\u03b5\u03bd \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b5 \u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03c1\u03bf\u03c6\u03b9\u03ba\u03ae \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03c0\u03b1\u03c1\u03ad\u03ba\u03b1\u03bc\u03c8\u03b5 \u03ad\u03bd\u03b1\u03bd \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03cc \u03c3\u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae; \u03a4\u03bf paper \u03b3\u03b9\u03b1 \u03c4\u03bf <strong>CRATE<\/strong> \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03b5\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae \u03c4\u03bf\u03c5 agent \u03c9\u03c2 \u03ad\u03bd\u03b1 \u03b1\u03b4\u03b9\u03b1\u03c6\u03b1\u03bd\u03ad\u03c2 \u03c0\u03b1\u03ba\u03ad\u03c4\u03bf.<\/p>\n<p>\u03a4\u03bf CRATE, \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b1\u03c1\u03c7\u03b9\u03ba\u03ac <em>Consequence Reasoning and Aggregation for Trajectory Evaluation<\/em>, \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf VLM-as-judge \u03b4\u03cd\u03bf \u03c3\u03c4\u03b1\u03b4\u03af\u03c9\u03bd. \u03a0\u03c1\u03ce\u03c4\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac, \u03b5\u03be\u03ac\u03b3\u03b5\u03b9 \u03c4\u03b1 \u03bf\u03c0\u03c4\u03b9\u03ba\u03ac \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1 \u03c0\u03bf\u03c5 \u03c3\u03c7\u03b5\u03c4\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u03c2. \u0388\u03c0\u03b5\u03b9\u03c4\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03bd\u03c4\u03c1\u03ce\u03bd\u03b5\u03b9 \u03b1\u03c5\u03c4\u03ac \u03c4\u03b1 \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03b1 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03b6\u03b5\u03b9 \u03b1\u03bd \u03b9\u03ba\u03b1\u03bd\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd \u03cc\u03bb\u03b5\u03c2 \u03bf\u03b9 \u03b1\u03c0\u03b1\u03b9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2. \u0397 \u03b5\u03c0\u03ad\u03ba\u03c4\u03b1\u03c3\u03b7 CRATE-S \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03c3\u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1.<\/p>\n<p>\u0393\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd <a href=\"https:\/\/twodots.gr\/ai-agents-paragogi-leitourgikes-astochies\/\">AI workflows \u03c0\u03bf\u03c5 \u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03b1\u03c0\u03cc demo \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/a>, customer-service agents \u03ae \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2, \u03c4\u03bf \u03b5\u03bd\u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03bd \u03b4\u03b5\u03bd \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b1 benchmarks Android. \u0397 \u03ad\u03c1\u03b5\u03c5\u03bd\u03b1 \u03b1\u03bd\u03b1\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b1\u03c1\u03c7\u03ae \u03b4\u03b9\u03b1\u03ba\u03c5\u03b2\u03ad\u03c1\u03bd\u03b7\u03c3\u03b7\u03c2: \u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1, \u03ba\u03b1\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1, \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03b9\u03bc\u03b1 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1.<\/p>\n<h2 id=\"teliki-othoni-den-apodeiknyei-diadromi\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03bb\u03b7 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae<\/h2>\n<p>\u039f\u03b9 \u03c0\u03c1\u03ce\u03c4\u03b5\u03c2 offline \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03b9 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2 \u03c3\u03c5\u03bd\u03ad\u03ba\u03c1\u03b9\u03bd\u03b1\u03bd \u03c4\u03b9\u03c2 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c8\u03b5\u03b9\u03c2 \u03c4\u03bf\u03c5 agent \u03bc\u03b5 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 ground truths. \u0395\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03ce\u03b3\u03b9\u03bc\u03b5\u03c2, \u03b1\u03bb\u03bb\u03ac \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c4\u03b9\u03bc\u03c9\u03c1\u03ae\u03c3\u03bf\u03c5\u03bd \u03bc\u03b9\u03b1 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae, \u03b5\u03be\u03af\u03c3\u03bf\u03c5 \u03ad\u03b3\u03ba\u03c5\u03c1\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae. \u039f\u03b9 online rule-based \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03b9 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b5\u03c2 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b2\u03ac\u03bb\u03bb\u03bf\u03bd, \u03b1\u03c0\u03b1\u03b9\u03c4\u03bf\u03cd\u03bd \u03cc\u03bc\u03c9\u03c2 \u03b5\u03ba\u03c4\u03b5\u03c4\u03b1\u03bc\u03ad\u03bd\u03b7 \u03c7\u03b5\u03b9\u03c1\u03bf\u03ba\u03af\u03bd\u03b7\u03c4\u03b7 \u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ae \u03ba\u03b1\u03bd\u03cc\u03bd\u03c9\u03bd \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 task \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03ba\u03bb\u03b9\u03bc\u03b1\u03ba\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b5\u03cd\u03ba\u03bf\u03bb\u03b1 \u03c3\u03b5 \u03bd\u03ad\u03b5\u03c2 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2.<\/p>\n<p>\u039f\u03b9 model-based evaluators \u03bc\u03b5 vision-language models \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7, \u03cc\u03bc\u03c9\u03c2 \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 \u03c6\u03bf\u03c1\u03c4\u03ce\u03bd\u03bf\u03c5\u03bd \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1 screenshots \u03ba\u03b1\u03b9 \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03b9\u03ce\u03bd \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1, \u03ad\u03bd\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bb\u03cd\u03bd\u03b5\u03b9 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/ai-agent-harnesses-elegchos-kostos-self-hosting\/\">agent harness<\/a>. \u03a4\u03bf paper \u03bf\u03bd\u03bf\u03bc\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 <em>context overload<\/em>: \u03bf evaluator \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03ae\u03c3\u03b5\u03b9 \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b5\u03c2, \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2 \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7\u03c2, \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c7\u03ad\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b5\u03c2 \u03c4\u03b7\u03c2 \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae\u03c2. \u03a3\u03c5\u03c7\u03bd\u03ac \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b9\u03b4\u03b9\u03cc\u03ba\u03c4\u03b7\u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03b5 \u03c0\u03bf\u03bb\u03cd \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf context window.<\/p>\n<p>\u0397 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03b1\u03bd\u03b5\u03c0\u03b1\u03c1\u03ba\u03ae\u03c2. \u039c\u03b9\u03b1 \u03bf\u03b8\u03cc\u03bd\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03ac\u03c1\u03b8\u03c1\u03bf \u03b1\u03bd\u03bf\u03b9\u03ba\u03c4\u03cc, \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03bc\u03b7 \u03b2\u03b5\u03b2\u03b1\u03b9\u03ce\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03bf agent \u03ac\u03bd\u03bf\u03b9\u03be\u03b5 \u03b4\u03cd\u03bf \u03ac\u03c1\u03b8\u03c1\u03b1 \u03cc\u03c0\u03c9\u03c2 \u03b6\u03b7\u03c4\u03ae\u03b8\u03b7\u03ba\u03b5. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1 \u03ad\u03c4\u03bf\u03b9\u03bc\u03bf, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03b1\u03bd \u03c3\u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae \u03b5\u03c0\u03b9\u03bb\u03ad\u03c7\u03b8\u03b7\u03ba\u03b5 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03c0\u03b1\u03c1\u03b1\u03bb\u03ae\u03c0\u03c4\u03b7\u03c2. \u0397 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03b4\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03b7\u03c2 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03c4\u03c9\u03bd \u03cc\u03c1\u03c9\u03bd \u03c4\u03bf\u03c5 task, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5 \u03c4\u03b5\u03bb\u03b5\u03c5\u03c4\u03b1\u03af\u03bf\u03c5 frame.<\/p>\n<div class=\"td-comparison td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid--three\">\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">\u039c\u03cc\u03bd\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7<\/p>\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03c6\u03b8\u03b7\u03bd\u03ae \u03c3\u03b5 context, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03c0\u03bf\u03c3\u03cc\u03c4\u03b7\u03c4\u03b5\u03c2, \u03c3\u03b5\u03b9\u03c1\u03ac \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03b9\u03ce\u03bd \u03ae \u03c0\u03c1\u03bf\u03c3\u03c9\u03c1\u03b9\u03bd\u03ae \u03ad\u03ba\u03b8\u03b5\u03c3\u03b7 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03c0\u03bf\u03c5 \u03c3\u03c5\u03bd\u03ad\u03b2\u03b7 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf \u03c4\u03b5\u03bb\u03b5\u03c5\u03c4\u03b1\u03af\u03bf frame.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">\u03a7\u03b1\u03bc\u03b7\u03bb\u03cc context<\/span><span class=\"td-badge\">\u03a5\u03c8\u03b7\u03bb\u03ae \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b1 evidence<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">\u039f\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 raw trajectory<\/p>\n<p>\u0394\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af screenshots \u03ba\u03b1\u03b9 actions, \u03cc\u03bc\u03c9\u03c2 \u03b1\u03bd\u03b1\u03b3\u03ba\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf\u03bd evaluator \u03bd\u03b1 \u03bb\u03cd\u03c3\u03b5\u03b9 \u03bf\u03c0\u03c4\u03b9\u03ba\u03cc \u03b8\u03cc\u03c1\u03c5\u03b2\u03bf, \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c7\u03ad\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03cc\u03bb\u03bf\u03c5\u03c2 \u03c4\u03bf\u03c5\u03c2 \u03cc\u03c1\u03bf\u03c5\u03c2 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03c6\u03bf\u03c1\u03c4\u03c9\u03bc\u03ad\u03bd\u03bf prompt.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">\u03a0\u03bb\u03ae\u03c1\u03b5\u03c2 \u03af\u03c7\u03bd\u03bf\u03c2<\/span><span class=\"td-badge\">Context overload<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">CRATE evidence trail<\/p>\n<p>\u03a3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 screen-action pair \u03c3\u03b5 task clue \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1, \u03ad\u03c0\u03b5\u03b9\u03c4\u03b1 \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03c1\u03b7\u03c4\u03ae \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b7. \u03a0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b5\u03be\u03b1\u03c1\u03c4\u03b7\u03bc\u03ad\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1 \u03c4\u03bf\u03c5 VLM.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Step-level evidence<\/span><span class=\"td-badge\">\u0395\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03c3\u03cd\u03bd\u03b8\u03b5\u03c3\u03b7<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"proto-stadio-synepeies-ana-vima\">\u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf: \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b5\u03c2 \u03b1\u03bd\u03ac \u03b2\u03ae\u03bc\u03b1 \u03b1\u03bd\u03c4\u03af \u03b3\u03b9\u03b1 \u03bf\u03c0\u03c4\u03b9\u03ba\u03cc \u03b8\u03cc\u03c1\u03c5\u03b2\u03bf<\/h2>\n<p>\u039c\u03b9\u03b1 trajectory \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae, \u03c4\u03b7 \u03c3\u03b5\u03b9\u03c1\u03ac \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03b9\u03ce\u03bd \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03c3\u03b5\u03b9\u03c1\u03ac screenshots. \u0393\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1, \u03c4\u03bf CRATE \u03b4\u03af\u03bd\u03b5\u03b9 \u03c3\u03c4\u03bf VLM \u03c4\u03b7\u03bd \u03bf\u03b8\u03cc\u03bd\u03b7 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c4\u03b7\u03bd \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c1\u03c7\u03b9\u03ba\u03ae \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1. \u03a4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03b4\u03cd\u03bf \u03b5\u03af\u03b4\u03b7 \u03ba\u03b5\u03b9\u03bc\u03ad\u03bd\u03bf\u03c5: \u03c4\u03b1 \u03bf\u03c1\u03b1\u03c4\u03ac \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1 \u03c0\u03bf\u03c5 \u03c3\u03c7\u03b5\u03c4\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf\u03bd \u03c3\u03c4\u03cc\u03c7\u03bf \u03ba\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03bf\u03c5\u03b4\u03ad\u03c4\u03b5\u03c1\u03b7 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03c4\u03bf\u03c5 \u03ac\u03bc\u03b5\u03c3\u03bf\u03c5 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u03c2.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03c4\u03b7\u03c2 \u03c0\u03c1\u03bf\u03b3\u03b5\u03bd\u03ad\u03c3\u03c4\u03b5\u03c1\u03b7\u03c2 \u03bf\u03b8\u03cc\u03bd\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u03c2, \u03b1\u03bd\u03c4\u03af \u03c4\u03bf\u03c5 \u03ba\u03bb\u03b1\u03c3\u03b9\u03ba\u03bf\u03cd triplet \u03c0\u03c1\u03b9\u03bd\u2013\u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u2013\u03bc\u03b5\u03c4\u03ac, \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b7\u03c0\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b5\u03be\u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd \u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03c9\u03bd. \u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b2\u03b1\u03c3\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd \u03b9\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 VLM \u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03b5\u03bb\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03ba\u03c5\u03c1\u03b9\u03bf\u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u03c2. \u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03ba\u03ac\u03b8\u03b5 \u03b6\u03b5\u03cd\u03b3\u03bf\u03c2 \u03b1\u03bd\u03b1\u03bb\u03cd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1, \u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 batched inference.<\/p>\n<p>\u03a4\u03bf \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 semantic compression. \u0386\u03c3\u03c7\u03b5\u03c4\u03b1 \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1 \u03c4\u03bf\u03c5 UI \u03c6\u03b9\u03bb\u03c4\u03c1\u03ac\u03c1\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03b7 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03bf\u03c0\u03c4\u03b9\u03ba\u03ae \u03c4\u03c1\u03bf\u03c7\u03b9\u03ac \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03bd. \u0394\u03b5\u03bd \u03b5\u03be\u03b1\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ae \u03b1\u03bb\u03c5\u03c3\u03af\u03b4\u03b1\u00b7 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b9\u03bf \u03c0\u03c5\u03ba\u03bd\u03ae \u03ba\u03b1\u03b9 \u03b5\u03c5\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b7 \u03c3\u03c4\u03b7\u03bd \u03b5\u03c0\u03b5\u03be\u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03ad\u03c4\u03c1\u03b9\u03bf\u03c5 \u03bc\u03b5\u03b3\u03ad\u03b8\u03bf\u03c5\u03c2.<\/p>\n<h2 id=\"deytero-stadio-elegchos-oron-ergasias\">\u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf: \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03cc\u03bb\u03c9\u03bd \u03c4\u03c9\u03bd \u03cc\u03c1\u03c9\u03bd \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2<\/h2>\n<p>\u03a3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03ad\u03bd\u03c4\u03c1\u03c9\u03c3\u03b7 \u03b5\u03c0\u03b9\u03c0\u03ad\u03b4\u03bf\u03c5 trajectory, \u03bf evaluator \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c1\u03c7\u03b9\u03ba\u03ae \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae, \u03cc\u03bb\u03b1 \u03c4\u03b1 step-level \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7. \u03a0\u03c1\u03ce\u03c4\u03b1 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03bd\u03b1\u03b3\u03ba\u03b1\u03af\u03b1 \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b7 \u03c0\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03b5\u03c4\u03b1\u03b9 \u03c1\u03b7\u03c4\u03ac \u03c3\u03c4\u03b7\u03bd \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae. \u039c\u03b5\u03c4\u03ac \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03bc\u03af\u03b1 \u03c0\u03c1\u03bf\u03c2 \u03bc\u03af\u03b1 \u03b1\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03ac\u03bc\u03b5\u03c3\u03bf, \u03c3\u03b1\u03c6\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03bc\u03b7 \u03b1\u03bc\u03c6\u03af\u03c3\u03b7\u03bc\u03bf \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf \u03cc\u03c4\u03b9 \u03b9\u03ba\u03b1\u03bd\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5.<\/p>\n<p>\u0397 \u03bf\u03b4\u03b7\u03b3\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bd\u03c4\u03b7\u03c1\u03b7\u03c4\u03b9\u03ba\u03ae: \u03b1\u03bd \u03ad\u03c3\u03c4\u03c9 \u03bc\u03af\u03b1 \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b7 \u03bb\u03b5\u03af\u03c0\u03b5\u03b9 \u03ae \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03c3\u03b1\u03c6\u03ae\u03c2, \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c7\u03b1\u03c1\u03b1\u03ba\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03b7\u03bc\u03ad\u03bd\u03b7. \u0391\u03c5\u03c4\u03cc \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03b1\u03b9\u03c3\u03b9\u03cc\u03b4\u03bf\u03be\u03b5\u03c2 \u03c5\u03c0\u03bf\u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03cc\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c0\u03c1\u03bf\u03c3\u03c0\u03b1\u03b8\u03b5\u03af \u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03ce\u03c3\u03b5\u03b9 \u03c4\u03b1 \u03ba\u03b5\u03bd\u03ac \u03bc\u03b9\u03b1\u03c2 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7\u03c2 \u03b1\u03ba\u03b1\u03c4\u03ad\u03c1\u03b3\u03b1\u03c3\u03c4\u03b7\u03c2 trajectory. \u0397 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03ba\u03c1\u03af\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03c5\u03b1\u03b4\u03b9\u03ba\u03ae, \u03b1\u03bb\u03bb\u03ac \u03c3\u03c5\u03bd\u03bf\u03b4\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03b1\u03b9\u03c4\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03b1\u03bc\u03bc\u03b9\u03c3\u03bc\u03ad\u03bd\u03b7 \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03b1\u03c0\u03b1\u03b9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2.<\/p>\n<p>\u03a3\u03b5 \u03ad\u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1 \u03c4\u03bf\u03c5 paper, \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b1\u03c0\u03b1\u03b9\u03c4\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03b1\u03bd\u03bf\u03af\u03be\u03bf\u03c5\u03bd \u03b4\u03cd\u03bf \u03ac\u03c1\u03b8\u03c1\u03b1. \u03a4\u03bf CRATE \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c1\u03b9\u03c3\u03b5 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b5\u03c0\u03b9\u03bc\u03ad\u03c1\u03bf\u03c5\u03c2 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03cc\u03c4\u03b9 \u03b1\u03bd\u03bf\u03af\u03c7\u03c4\u03b7\u03ba\u03b5 \u03bc\u03cc\u03bd\u03bf \u03ad\u03bd\u03b1, \u03b5\u03bd\u03ce \u03c4\u03bf SPA-Bench \u03b4\u03b5\u03bd \u03ad\u03bb\u03b1\u03b2\u03b5 \u03c3\u03c9\u03c3\u03c4\u03ac \u03c5\u03c0\u03cc\u03c8\u03b7 \u03c4\u03bf\u03bd \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03cc \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03cc \u03ba\u03b1\u03b9 \u03ad\u03b4\u03c9\u03c3\u03b5 false positive. \u03a4\u03bf \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 condition coverage \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03bb\u03ae \u03bf\u03c0\u03c4\u03b9\u03ba\u03ae \u03bf\u03bc\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bc\u03b5 \u03ad\u03bd\u03b1\u03bd \u03b1\u03bd\u03b1\u03bc\u03b5\u03bd\u03cc\u03bc\u03b5\u03bd\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc \u03c3\u03c4\u03cc\u03c7\u03bf.<\/p>\n<h2 id=\"crate-s-asfaleia-kai-oloklirosi\">CRATE-S: \u03b7 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03b4\u03b5\u03bd \u03c4\u03b1\u03c5\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7<\/h2>\n<p>\u0388\u03bd\u03b1\u03c2 agent \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03ce\u03c3\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac \u03bd\u03b1 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9 \u03b5\u03c0\u03b9\u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03b1, \u03b3\u03b9\u2019 \u03b1\u03c5\u03c4\u03cc \u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03ad\u03bd\u03b1 <a href=\"https:\/\/twodots.gr\/agentic-ai-security-stack-controls-technologikes-ependyseis\/\">agentic AI security stack<\/a>. \u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03cc\u03c0\u03c9\u03c2 \u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03c1\u03bf\u03c6\u03b9\u03ba\u03ad\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2, \u03b4\u03b9\u03b1\u03c1\u03c1\u03bf\u03ae \u03b9\u03b4\u03b9\u03c9\u03c4\u03b9\u03ba\u03ce\u03bd \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03b9\u03ce\u03bd \u03ae \u03b4\u03b9\u03ac\u03b4\u03bf\u03c3\u03b7 \u03b5\u03c0\u03b9\u03b2\u03bb\u03b1\u03b2\u03bf\u03cd\u03c2 \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03bf\u03bc\u03ad\u03bd\u03bf\u03c5. \u03a4\u03bf CRATE-S \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 \u03b1\u03c5\u03c4\u03ad\u03c2 \u03c4\u03b9\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b5\u03c2 \u03c4\u03cc\u03c3\u03bf \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1 \u03cc\u03c3\u03bf \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c3\u03cd\u03bd\u03bf\u03bb\u03bf \u03c4\u03b7\u03c2 trajectory.<\/p>\n<p>\u03a3\u03c4\u03bf safety mode \u03c0\u03b1\u03c1\u03b1\u03bb\u03b5\u03af\u03c0\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03c4\u03bf\u03c5 task, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03bf \u03c3\u03c4\u03cc\u03c7\u03bf\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 task-agnostic \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac\u03c2 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03ba\u03c1\u03af\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c4\u03bf \u03b1\u03bd \u03b7 \u03af\u03b4\u03b9\u03b1 \u03b7 \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae \u03ae\u03c4\u03b1\u03bd \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae\u03c2. \u039a\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03bf\u03c0\u03c4\u03b9\u03ba\u03ac clues, \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03b4\u03c5\u03b1\u03b4\u03b9\u03ba\u03cc risk label. \u0397 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03c3\u03c5\u03b3\u03ba\u03ad\u03bd\u03c4\u03c1\u03c9\u03c3\u03b7 \u03b1\u03bd\u03b1\u03b6\u03b7\u03c4\u03ac \u03c1\u03b7\u03c4\u03ac \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03b9\u03bc\u03bf\u03c5\u03c2 \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5\u03c2 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7 \u03c9\u03c2 \u03bc\u03bf\u03bd\u03b1\u03b4\u03b9\u03ba\u03cc \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03bf.<\/p>\n<p>\u0397 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ae\u03c2 \u03b5\u03c0\u03b9\u03c3\u03ae\u03bc\u03b1\u03bd\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03b1\u03be\u03af\u03b1. \u0394\u03b5\u03bd \u03bb\u03ad\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03cc\u03c4\u03b9 \u03bc\u03b9\u03b1 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03ae\u03c4\u03b1\u03bd \u03b5\u03c0\u03b9\u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03b7, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf\u03c0\u03bf\u03b8\u03b5\u03c4\u03b5\u03af \u03c4\u03bf\u03bd \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03cc\u03c0\u03bf\u03c5 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b5. \u0391\u03c5\u03c4\u03cc \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b2\u03bf\u03b7\u03b8\u03ae\u03c3\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03bd\u03b1 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03c3\u03b5\u03b9 \u03b1\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 permission gate, \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7 \u03ad\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03ae \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae\u03c2 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03b9\u03ce\u03bd.<\/p>\n<aside class=\"td-article-note\">\n<p><strong>\u03a4\u03b1 step logs \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03b1\u03bb\u03bb\u03ac \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c6\u03ac\u03bd\u03b5\u03b9\u03b1 \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5.<\/strong> Screenshots, \u03c0\u03b1\u03c1\u03b1\u03bb\u03ae\u03c0\u03c4\u03b5\u03c2, \u03bc\u03b7\u03bd\u03cd\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 account state \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03bf\u03c5\u03bd \u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b9\u03ba\u03ac \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1. \u0388\u03bd\u03b1 production \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b5\u03bb\u03b1\u03c7\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03ae\u03c2, masking, \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7, \u03c7\u03c1\u03cc\u03bd\u03bf \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf review \u03bc\u03cc\u03bd\u03bf \u03b3\u03b9\u03b1 \u03c4\u03b9\u03c2 \u03b1\u03bc\u03c6\u03af\u03b2\u03bf\u03bb\u03b5\u03c2 \u03ae \u03c5\u03c8\u03b7\u03bb\u03bf\u03cd \u03c1\u03af\u03c3\u03ba\u03bf\u03c5 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ad\u03c2.<\/p>\n<\/aside>\n<h2 id=\"benchmarks-pragmatikes-syskeves-etiketes\">\u03a4\u03c1\u03af\u03b1 benchmarks, \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c5\u03c3\u03ba\u03b5\u03c5\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03c1\u03b7\u03c4\u03ad\u03c2 \u03b5\u03c4\u03b9\u03ba\u03ad\u03c4\u03b5\u03c2<\/h2>\n<p>\u0393\u03b9\u03b1 \u03c4\u03bf task completion \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5 \u03c4\u03bf AndroidWorld \u03bc\u03b5 116 trajectories \u03b1\u03c0\u03cc M3A agent \u03c0\u03bf\u03c5 \u03b2\u03b1\u03c3\u03b9\u03b6\u03cc\u03c4\u03b1\u03bd \u03c3\u03c4\u03bf Qwen2.5-VL-72B-Instruct. \u039f\u03b9 rule-based \u03b5\u03c4\u03b9\u03ba\u03ad\u03c4\u03b5\u03c2 \u03c0\u03b5\u03c1\u03b9\u03bb\u03ac\u03bc\u03b2\u03b1\u03bd\u03b1\u03bd 27 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b5\u03c2 \u03ba\u03b1\u03b9 89 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b5\u03c2. \u0393\u03b9\u03b1 \u03c4\u03b7\u03bd \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5 \u03c4\u03bf MobileRisk \u03bc\u03b5 102 \u03b1\u03c3\u03c6\u03b1\u03bb\u03b5\u03af\u03c2 \u03ba\u03b1\u03b9 102 \u03bc\u03b7 \u03b1\u03c3\u03c6\u03b1\u03bb\u03b5\u03af\u03c2 trajectories \u03b1\u03c0\u03cc mobile agent \u03b2\u03b1\u03c3\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf \u03c3\u03c4\u03bf GPT-4o, \u03bc\u03b5 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7 \u03b5\u03c0\u03b9\u03c3\u03ae\u03bc\u03b1\u03bd\u03c3\u03b7 \u03c3\u03b5 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf \u03b2\u03ae\u03bc\u03b1\u03c4\u03bf\u03c2 \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae\u03c2 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae\u03c2.<\/p>\n<p>\u039f\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03b4\u03b7\u03bc\u03b9\u03bf\u03cd\u03c1\u03b3\u03b7\u03c3\u03b1\u03bd \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03c4\u03bf CRATEBench: 187 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b5\u03c2 \u03c3\u03b5 35 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03ad\u03be\u03b9 domains. \u039f\u03b9 24 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2 \u03ae\u03c4\u03b1\u03bd \u03b4\u03b7\u03bc\u03bf\u03c6\u03b9\u03bb\u03b5\u03af\u03c2 \u03c3\u03c4\u03b7\u03bd \u039a\u03af\u03bd\u03b1 \u03ba\u03b1\u03b9 \u03bf\u03b9 11 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03bd \u03b4\u03b9\u03b5\u03b8\u03bd\u03ce\u03c2. \u03a0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03bf 65% \u03c4\u03c9\u03bd tasks \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03c3\u03c4\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc \u03b5\u03b9\u03b4\u03b9\u03ba\u03bf\u03cd\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03c1\u03b5\u03b9\u03c2 \u03b3\u03cd\u03c1\u03bf\u03c5\u03c2 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 refinement, \u03b5\u03bd\u03ce \u03c4\u03b1 \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03b1 \u03c0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03cc\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd \u03b1\u03c0\u03cc \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bc\u03b5\u03bd\u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03c5\u03c0\u03bf\u03bf\u03bc\u03ac\u03b4\u03b1 62 tasks, \u03c4\u03bf UI-TARS-7B \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b5 \u03c4\u03b9\u03c2 \u03b5\u03bd\u03c4\u03bf\u03bb\u03ad\u03c2 \u03c3\u03b5 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03ba\u03b9\u03bd\u03b7\u03c4\u03ae \u03c3\u03c5\u03c3\u03ba\u03b5\u03c5\u03ae, \u03ce\u03c3\u03c4\u03b5 \u03bf\u03b9 trajectories \u03bd\u03b1 \u03c0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03bf\u03c5\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac patterns \u03b1\u03c0\u03cc rendering, latency \u03ba\u03b1\u03b9 \u03b4\u03c5\u03bd\u03b1\u03bc\u03b9\u03ba\u03cc UI. \u0397 \u03c7\u03b5\u03b9\u03c1\u03bf\u03ba\u03af\u03bd\u03b7\u03c4\u03b7 \u03b5\u03c0\u03b1\u03bb\u03ae\u03b8\u03b5\u03c5\u03c3\u03b7 \u03ad\u03b4\u03c9\u03c3\u03b5 24 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b5\u03c2 \u03ba\u03b1\u03b9 38 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b5\u03c2. \u0388\u03c4\u03c3\u03b9 \u03c4\u03bf \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03b4\u03b5\u03bd \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b5 \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 emulator \u03ae \u03c4\u03b5\u03c7\u03bd\u03b7\u03c4\u03ad\u03c2 \u03c3\u03c4\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">\u03a4\u03bf \u03b5\u03cd\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2 \u03c3\u03b5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03cd\u03c2<\/p>\n<p class=\"td-chart-subtitle\">\u039f\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b1 datasets \u03c4\u03b7\u03c2 \u03c0\u03c1\u03bf\u03b4\u03b7\u03bc\u03bf\u03c3\u03af\u03b5\u03c5\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03bf\u03c5\u03bd \u03c4\u03bf benchmark, \u03cc\u03c7\u03b9 \u03b1\u03bd\u03b1\u03bc\u03b5\u03bd\u03cc\u03bc\u03b5\u03bd\u03b7 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">116<\/span><span class=\"td-metric-label\">AndroidWorld trajectories \u03b3\u03b9\u03b1 task completion<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">204<\/span><span class=\"td-metric-label\">MobileRisk trajectories, 102 safe \u03ba\u03b1\u03b9 102 unsafe<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">187<\/span><span class=\"td-metric-label\">CRATEBench tasks \u03c3\u03b5 35 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03ad\u03be\u03b9 domains<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">62<\/span><span class=\"td-metric-label\">tasks \u03b5\u03ba\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ad\u03bd\u03b1 \u03c3\u03b5 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03ba\u03b9\u03bd\u03b7\u03c4\u03ae \u03c3\u03c5\u03c3\u03ba\u03b5\u03c5\u03ae<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<aside class=\"td-article-note\">\n<p><strong>\u03a0\u03c1\u03cc\u03ba\u03b5\u03b9\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c0\u03c1\u03bf\u03b4\u03b7\u03bc\u03bf\u03c3\u03af\u03b5\u03c5\u03c3\u03b7 v1 \u03c4\u03b7\u03c2 21\u03b7\u03c2 \u0391\u03c5\u03b3\u03bf\u03cd\u03c3\u03c4\u03bf\u03c5 2026.<\/strong> \u03a4\u03b1 scores \u03bc\u03b5\u03c4\u03c1\u03bf\u03cd\u03bd \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03af\u03b1 \u03bc\u03b5 rule-based \u03ae \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 ground-truth labels \u03c3\u03c4\u03b1 AndroidWorld, CRATEBench \u03ba\u03b1\u03b9 MobileRisk. \u0394\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b7 \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1\u03c2, \u03b3\u03b5\u03bd\u03af\u03ba\u03b5\u03c5\u03c3\u03b7 \u03c3\u03b5 web\/desktop agents \u03ae \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 real-time \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1\u03c2.<\/p>\n<\/aside>\n<h2 id=\"apotelesmata-task-completion\">\u03a4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c3\u03c4\u03bf task completion<\/h2>\n<p>\u039c\u03b5 Qwen2.5-VL \u03c9\u03c2 evaluator, \u03c4\u03bf CRATE \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03c3\u03c4\u03bf AndroidWorld accuracy 0,930 \u03ba\u03b1\u03b9 F1-score 0,833. \u03a3\u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c1\u03cd\u03b8\u03bc\u03b9\u03c3\u03b7, \u03c4\u03bf SPA-Bench \u03b5\u03af\u03c7\u03b5 F1 0,623, \u03c4\u03bf A3 \u03bc\u03b5 final-state evaluation 0,773 \u03ba\u03b1\u03b9 \u03c4\u03bf A3 \u03bc\u03b5 essential-state evaluation 0,638. \u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03b7 \u03c9\u03c2 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 20% \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 SPA-Bench.<\/p>\n<p>\u03a3\u03c4\u03bf CRATEBench \u03bc\u03b5 Qwen2.5-VL, \u03c4\u03bf CRATE \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 accuracy 0,839 \u03ba\u03b1\u03b9 F1 0,815. \u039c\u03b5 GPT-4o evaluator, \u03bf\u03b9 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03c2 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03ae\u03c4\u03b1\u03bd 0,806 \u03ba\u03b1\u03b9 0,760. \u03a4\u03bf CRATE \u03b5\u03af\u03c7\u03b5 \u03c4\u03b7\u03bd \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03b7 accuracy \u03ba\u03b1\u03b9 \u03c4\u03bf \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf F1 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd \u03c4\u03c9\u03bd \u03c3\u03c5\u03b3\u03ba\u03c1\u03b9\u03bd\u03cc\u03bc\u03b5\u03bd\u03c9\u03bd \u03bc\u03b5\u03b8\u03cc\u03b4\u03c9\u03bd \u03ba\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b1 \u03b4\u03cd\u03bf evaluator backbones \u03c3\u03c4\u03b1 \u03b4\u03cd\u03bf task-completion benchmarks.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae evaluator \u03b5\u03c0\u03b7\u03c1\u03ad\u03b1\u03c3\u03b5 \u03c4\u03bf \u03c0\u03c1\u03bf\u03c6\u03af\u03bb \u03c3\u03c6\u03b1\u03bb\u03bc\u03ac\u03c4\u03c9\u03bd. \u03a3\u03c4\u03b9\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03c1\u03c5\u03b8\u03bc\u03af\u03c3\u03b5\u03b9\u03c2 \u03c4\u03bf GPT-4o \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c0\u03b9\u03bf \u03c3\u03c5\u03bd\u03c4\u03b7\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc, \u03bc\u03b5 \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03b7 precision \u03ba\u03b1\u03b9 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03b7 recall. \u03a4\u03bf Qwen2.5-VL \u03b1\u03c0\u03ad\u03b4\u03c9\u03c3\u03b5 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 \u03cc\u03c4\u03b1\u03bd \u03bb\u03ac\u03bc\u03b2\u03b1\u03bd\u03b5 \u03c4\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03b1 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03c4\u03bf\u03c5 CRATE, \u03b5\u03bd\u03ce \u03c4\u03bf GPT-4o \u03b2\u03bf\u03ae\u03b8\u03b7\u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03c4\u03bf SPA-Bench \u03c0\u03bf\u03c5 \u03b4\u03b9\u03b1\u03c7\u03b5\u03b9\u03c1\u03b9\u03b6\u03cc\u03c4\u03b1\u03bd \u03c0\u03b9\u03bf \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03bf \u03ba\u03b1\u03b9 \u03b5\u03bc\u03c0\u03bb\u03b5\u03b3\u03bc\u03ad\u03bd\u03bf context. \u0391\u03c5\u03c4\u03cc \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03af \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1 \u03c4\u03c9\u03bd \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03c9\u03bd \u03c0\u03b5\u03b9\u03c1\u03b1\u03bc\u03ac\u03c4\u03c9\u03bd, \u03cc\u03c7\u03b9 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7 \u03c4\u03c9\u03bd \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd.<\/p>\n<h2 id=\"dokimes-operational-safety\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2 operational safety<\/h2>\n<p>\u03a3\u03c4\u03bf MobileRisk, \u03c4\u03bf CRATE-S \u03bc\u03b5 Qwen2.5-VL \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 trajectory-level accuracy 0,711, F1 0,697, precision 0,731 \u03ba\u03b1\u03b9 recall 0,667. \u039f step-level score \u03ae\u03c4\u03b1\u03bd 28,8, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 23,5 \u03b3\u03b9\u03b1 \u03c4\u03bf OS-Sentinel. \u039c\u03b5 GPT-4o, \u03c4\u03bf CRATE-S \u03b5\u03af\u03c7\u03b5 F1 0,702 \u03ba\u03b1\u03b9 recall 0,902, \u03b5\u03bd\u03ce \u03bf step-level score \u03ae\u03c4\u03b1\u03bd 27,1 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 25,7 \u03c4\u03bf\u03c5 OS-Sentinel.<\/p>\n<p>\u0397 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ae VLM-only \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03bc\u03b5 Qwen2.5-VL \u03b5\u03af\u03c7\u03b5 recall \u03bc\u03cc\u03bb\u03b9\u03c2 0,245, \u03b5\u03bd\u03ce \u03bf rule-based evaluator \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03b6\u03cc\u03c4\u03b1\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf\u03c5\u03c2 \u03ac\u03ba\u03b1\u03bc\u03c0\u03c4\u03bf\u03c5\u03c2 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf\u03c5\u03c2 \u03ba\u03b1\u03bd\u03cc\u03bd\u03b5\u03c2. \u03a4\u03bf OS-Sentinel \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03bd\u03b5 \u03c4\u03b7 recall \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03c2 rule-based \u03ba\u03b1\u03b9 VLM \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5 \u03bb\u03bf\u03b3\u03b9\u03ba\u03cc OR, \u03b1\u03bb\u03bb\u03ac \u03b8\u03c5\u03c3\u03af\u03b1\u03b6\u03b5 precision. \u03a4\u03bf CRATE-S \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b5 \u03bd\u03b1 \u03c0\u03b5\u03c4\u03cd\u03c7\u03b5\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b9\u03c3\u03bf\u03c1\u03c1\u03bf\u03c0\u03af\u03b1 \u03bc\u03ad\u03c3\u03c9 \u03c1\u03b7\u03c4\u03ae\u03c2 \u03b1\u03b9\u03c4\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1\u03c2 \u03ba\u03ac\u03b8\u03b5 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u03c2.<\/p>\n<p>\u0397 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03c7\u03c1\u03cc\u03bd\u03bf\u03c5 \u03b5\u03bd\u03c4\u03bf\u03c0\u03b9\u03c3\u03bc\u03bf\u03cd \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af penalty window \u03c0\u03ad\u03bd\u03c4\u03b5 \u03b2\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd: \u03b7 \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03b3\u03b9\u03c3\u03c4\u03b7 \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03bc\u03b7 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 \u03b2\u03ae\u03bc\u03b1 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b3\u03c1\u03b1\u03bc\u03bc\u03b9\u03ba\u03ac \u03cc\u03c3\u03bf \u03b7 \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7 \u03b1\u03c0\u03bf\u03bc\u03b1\u03ba\u03c1\u03cd\u03bd\u03b5\u03c4\u03b1\u03b9. \u0386\u03c1\u03b1 \u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03c0\u03b9\u03b2\u03c1\u03b1\u03b2\u03b5\u03cd\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03b5\u03c4\u03b9\u03ba\u03ad\u03c4\u03b1, \u03b1\u03bb\u03bb\u03ac \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03ad\u03b3\u03ba\u03b1\u03b9\u03c1\u03b7 \u03c4\u03bf\u03c0\u03bf\u03b8\u03ad\u03c4\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5.<\/p>\n<h2 id=\"ablations-simasia-sympiesis\">\u039f\u03b9 ablations \u03b5\u03be\u03b7\u03b3\u03bf\u03cd\u03bd \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1<\/h2>\n<p>\u03a3\u03c4\u03bf CRATEBench, \u03b7 \u03c0\u03bb\u03ae\u03c1\u03b7\u03c2 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03bc\u03b5 Qwen2.5-VL \u03b5\u03af\u03c7\u03b5 F1 0,815. \u03a7\u03c9\u03c1\u03af\u03c2 step-level reasoning, \u03cc\u03c4\u03b1\u03bd \u03bf\u03b9 \u03b1\u03ba\u03b1\u03c4\u03ad\u03c1\u03b3\u03b1\u03c3\u03c4\u03b5\u03c2 \u03bf\u03b8\u03cc\u03bd\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03c3\u03b1\u03bd \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03c3\u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03c3\u03c5\u03b3\u03ba\u03ad\u03bd\u03c4\u03c1\u03c9\u03c3\u03b7, \u03c4\u03bf F1 \u03ad\u03c0\u03b5\u03c3\u03b5 \u03c3\u03b5 0,649. \u0397 recall \u03ad\u03b3\u03b9\u03bd\u03b5 1,000 \u03b1\u03bb\u03bb\u03ac \u03b7 precision \u03ad\u03c0\u03b5\u03c3\u03b5 \u03c3\u03b5 0,480: \u03ad\u03bd\u03b1 \u03bc\u03bf\u03c4\u03af\u03b2\u03bf \u03c5\u03c0\u03b5\u03c1\u03b2\u03bf\u03bb\u03b9\u03ba\u03ac \u03b1\u03b9\u03c3\u03b9\u03cc\u03b4\u03bf\u03be\u03c9\u03bd \u03ba\u03c1\u03af\u03c3\u03b5\u03c9\u03bd \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03c5\u03c0\u03b5\u03c1\u03c6\u03bf\u03c1\u03c4\u03c9\u03bc\u03ad\u03bd\u03bf context.<\/p>\n<p>\u038c\u03c4\u03b1\u03bd \u03b1\u03c6\u03b1\u03b9\u03c1\u03ad\u03b8\u03b7\u03ba\u03b5 \u03ba\u03ac\u03b8\u03b5 step-level \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03ae\u03b8\u03b7\u03ba\u03b5 \u03bc\u03cc\u03bd\u03bf \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7, \u03c4\u03bf Qwen2.5-VL \u03b1\u03bd\u03ad\u03ba\u03c4\u03b7\u03c3\u03b5 accuracy 0,822, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf F1 \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c3\u03c4\u03bf 0,731. \u0397 \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03b8\u03bf\u03c1\u03cd\u03b2\u03bf\u03c5 \u03b2\u03bf\u03ae\u03b8\u03b7\u03c3\u03b5 \u03c4\u03b7\u03bd precision, \u03cc\u03bc\u03c9\u03c2 \u03c7\u03ac\u03b8\u03b7\u03ba\u03b1\u03bd \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b5\u03c2 \u03b3\u03b9\u03b1 \u03c4\u03b7 \u03b4\u03c5\u03bd\u03b1\u03bc\u03b9\u03ba\u03ae \u03c4\u03b7\u03c2 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7\u03c2. \u039c\u03b5 GPT-4o, \u03c4\u03bf final-state-only F1 \u03ae\u03c4\u03b1\u03bd 0,611 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 0,760 \u03c4\u03b7\u03c2 \u03c0\u03bb\u03ae\u03c1\u03bf\u03c5\u03c2 \u03bc\u03b5\u03b8\u03cc\u03b4\u03bf\u03c5.<\/p>\n<p>\u03a4\u03bf \u03b4\u03af\u03b4\u03b1\u03b3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ce\u03c2 \u00ab\u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03bf context\u00bb. \u0397 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03b7\u03bd \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03ba\u03b1\u03b9 \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ae \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7. \u0388\u03bd\u03b1 \u03c0\u03bf\u03bb\u03cd \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf raw context \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03c0\u03b5\u03c1\u03b4\u03ad\u03c8\u03b5\u03b9 \u03c4\u03bf\u03bd evaluator, \u03b5\u03bd\u03ce \u03ad\u03bd\u03b1 \u03c5\u03c0\u03b5\u03c1\u03b2\u03bf\u03bb\u03b9\u03ba\u03ac \u03bc\u03b9\u03ba\u03c1\u03cc snapshot \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03af \u03c4\u03b9\u03c2 \u03b1\u03c0\u03bf\u03b4\u03b5\u03af\u03be\u03b5\u03b9\u03c2. \u03a4\u03bf CRATE \u03b1\u03bd\u03b1\u03b6\u03b7\u03c4\u03ac \u03c4\u03bf \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf: \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03b1 \u03b1\u03bb\u03bb\u03ac \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 step records.<\/p>\n<h2 id=\"business-control-layer\">\u0391\u03c0\u03cc \u03c4\u03bf research benchmark \u03c3\u03b5 \u03ad\u03bd\u03b1 business control layer<\/h2>\n<p>\u039c\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b5\u03af\u03b1 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af agents \u03c3\u03b5 CRM, \u03b7\u03bb\u03b5\u03ba\u03c4\u03c1\u03bf\u03bd\u03b9\u03ba\u03cc \u03b5\u03bc\u03c0\u03cc\u03c1\u03b9\u03bf \u03ae customer support \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03c1\u03ac\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c3\u03b5 \u03c4\u03c1\u03af\u03b1 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03b1 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03c4\u03b1 \u03b5\u03bd\u03c4\u03ac\u03be\u03b5\u03b9 \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03b5\u03c5\u03c1\u03cd\u03c4\u03b5\u03c1\u03bf \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf <a href=\"https:\/\/twodots.gr\/ai-agents-times-behavioral-certification\/\">behavioral certification<\/a>. \u03a0\u03c1\u03ce\u03c4\u03bf\u03bd, \u03bd\u03b1 \u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03c4\u03b9\u03c2 \u03c1\u03b7\u03c4\u03ad\u03c2 \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b5\u03c2 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1\u03c2. \u0394\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf\u03bd, \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 tool call \u03ae UI action \u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bc\u03b5\u03bd\u03b7 \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7, \u03c4\u03b7\u03bd \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03ac \u03c4\u03b7\u03c2. \u03a4\u03c1\u03af\u03c4\u03bf\u03bd, \u03bd\u03b1 \u03b6\u03b7\u03c4\u03ac \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03ba\u03c1\u03af\u03c3\u03b7 \u03bc\u03cc\u03bd\u03bf \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03b1\u03c5\u03c4\u03ac \u03c4\u03b1 \u03b5\u03c0\u03b9\u03bb\u03b5\u03b3\u03bc\u03ad\u03bd\u03b1 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc state.<\/p>\n<p>\u0393\u03b9\u03b1 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf\u03c5 \u03c1\u03af\u03c3\u03ba\u03bf\u03c5, \u03c4\u03bf \u03af\u03b4\u03b9\u03bf log \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c4\u03c1\u03bf\u03c6\u03bf\u03b4\u03bf\u03c4\u03b5\u03af \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc safety evaluator. \u0397 \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b7 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2: \u00ab\u03ad\u03b3\u03b9\u03bd\u03b5 \u03b1\u03c5\u03c4\u03cc \u03c0\u03bf\u03c5 \u03b6\u03b7\u03c4\u03ae\u03b8\u03b7\u03ba\u03b5;\u00bb \u03ba\u03b1\u03b9 \u00ab\u03ad\u03b3\u03b9\u03bd\u03b5 \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c0\u03b1\u03b3\u03bf\u03c1\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03ae \u03b5\u03c0\u03b9\u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03b7 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1;\u00bb. \u0397 \u03bc\u03af\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03b4\u03b5\u03bd \u03c3\u03c5\u03bd\u03b5\u03c0\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03c4\u03b7\u03bd \u03ac\u03bb\u03bb\u03b7.<\/p>\n<p>\u039f\u03b9 \u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03ad\u03c2 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03bd\u03b1 \u03be\u03b5\u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc success rate, \u03cc\u03c0\u03c9\u03c2 \u03c3\u03c5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03c9\u03bc\u03ad\u03bd\u03bf \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1 <a href=\"https:\/\/twodots.gr\/ai-agents-behavioral-testing\/\">behavioral testing \u03b3\u03b9\u03b1 AI agents<\/a>. Precision, recall, false positives, false negatives \u03ba\u03b1\u03b9 \u03c7\u03c1\u03cc\u03bd\u03bf\u03c2 \u03b5\u03bd\u03c4\u03bf\u03c0\u03b9\u03c3\u03bc\u03bf\u03cd \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5 \u03ad\u03c7\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u03a3\u03b5 \u03bc\u03b9\u03b1 \u03c1\u03bf\u03ae \u03cc\u03c0\u03bf\u03c5 \u03bc\u03b9\u03b1 \u03c8\u03b5\u03c5\u03b4\u03ae\u03c2 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bf\u03b4\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9 \u03c3\u03b5 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03b1\u03c0\u03bf\u03c3\u03c4\u03bf\u03bb\u03ae \u03ae \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd, \u03b7 conservative policy \u03c4\u03bf\u03c5 CRATE \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03b4\u03b9\u03b1\u03af\u03c4\u03b5\u03c1\u03b1 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-band-kicker\">\u039a\u03b1\u03bd\u03cc\u03bd\u03b1\u03c2 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2<\/p>\n<h3>\u03a0\u03cc\u03c4\u03b5 \u03bf mobile AI agent \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b7\u03bb\u03ce\u03c3\u03b5\u03b9 \u00ab\u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03ce\u03b8\u03b7\u03ba\u03b5\u00bb<\/h3>\n<p>\u039c\u03cc\u03bd\u03bf \u03cc\u03c4\u03b1\u03bd \u03ba\u03ac\u03b8\u03b5 \u03c1\u03b7\u03c4\u03ae \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b7 \u03c4\u03bf\u03c5 task \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf step evidence, \u03bf \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc\u03c2 safety evaluator \u03b4\u03b5\u03bd \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03b1\u03c0\u03b1\u03b3\u03bf\u03c1\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03c4\u03b1 false positives \u03ba\u03b1\u03b9 false negatives \u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03b5\u03bd\u03c4\u03cc\u03c2 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03b7\u03bc\u03ad\u03bd\u03c9\u03bd \u03bf\u03c1\u03af\u03c9\u03bd \u03ba\u03b1\u03b9 \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 human approval \u03b3\u03b9\u03b1 irreversible \u03ae sensitive actions. \u0391\u03bc\u03c6\u03af\u03b2\u03bf\u03bb\u03bf evidence \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 fail, handoff \u03ae \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 rollback.<\/p>\n<\/div>\n<h2 id=\"pilot-elegchou-mobile-ai-agent\">\u03a0\u03ce\u03c2 \u03c3\u03c4\u03ae\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 pilot \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2<\/h2>\n<p>\u03a4\u03bf CRATE \u03b5\u03af\u03bd\u03b1\u03b9 research framework, \u03cc\u03c7\u03b9 \u03ad\u03c4\u03bf\u03b9\u03bc\u03b7 \u03ac\u03b4\u03b5\u03b9\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1\u03c2. \u0388\u03bd\u03b1 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf pilot \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c4\u03b7 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03c4\u03bf\u03c5 \u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf workflow, \u03ba\u03c1\u03b1\u03c4\u03ac \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac success \u03ba\u03b1\u03b9 safety labels \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b7 \u03ba\u03c1\u03af\u03c3\u03b7 \u03bc\u03b5 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7 \u03b1\u03bd\u03b1\u03c6\u03bf\u03c1\u03ac \u03c0\u03c1\u03b9\u03bd \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03c3\u03b5\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03bf\u03cd\u03c2 \u03bb\u03bf\u03b3\u03b1\u03c1\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd\u03c2 \u03ae \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0391\u03c0\u03cc \u03c4\u03bf trajectory log \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf production gate<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Step 1<\/span><strong>\u0395\u03c0\u03b9\u03bb\u03ad\u03be\u03c4\u03b5 \u03ad\u03bd\u03b1 bounded workflow<\/strong>\n<p>\u039e\u03b5\u03ba\u03b9\u03bd\u03ae\u03c3\u03c4\u03b5 \u03bc\u03b5 \u03bc\u03af\u03b1 \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b1\u03bc\u03b2\u03b1\u03bd\u03cc\u03bc\u03b5\u03bd\u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c7\u03b1\u03bc\u03b7\u03bb\u03bf\u03cd \u03c1\u03af\u03c3\u03ba\u03bf\u03c5, \u03cc\u03c0\u03c9\u03c2 \u03c4\u03b1\u03be\u03b9\u03bd\u03cc\u03bc\u03b7\u03c3\u03b7 \u03b1\u03b9\u03c4\u03ae\u03bc\u03b1\u03c4\u03bf\u03c2 \u03ae \u03c0\u03c1\u03bf\u03b5\u03c4\u03bf\u03b9\u03bc\u03b1\u03c3\u03af\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2. \u0393\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/6-ai-agents-customer-support-gia-epicheiriseis\/\">customer-support agents<\/a>, \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c4\u03b7\u03bd \u03b1\u03c0\u03bf\u03c3\u03c4\u03bf\u03bb\u03ae \u03ba\u03b1\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03bb\u03bf\u03b3\u03b1\u03c1\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd \u03c0\u03af\u03c3\u03c9 \u03b1\u03c0\u03cc \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf gate.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 2<\/span><strong>\u0393\u03c1\u03ac\u03c8\u03c4\u03b5 condition contract<\/strong>\n<p>\u039c\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03c4\u03b7\u03bd \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b5\u03c2 \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b5\u03c2: \u03c3\u03c9\u03c3\u03c4\u03cc \u03b1\u03bd\u03c4\u03b9\u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf, \u03c3\u03c9\u03c3\u03c4\u03ae \u03c0\u03bf\u03c3\u03cc\u03c4\u03b7\u03c4\u03b1, \u03c3\u03c9\u03c3\u03c4\u03ae \u03c3\u03b5\u03b9\u03c1\u03ac, \u03b5\u03c0\u03b9\u03b2\u03b5\u03b2\u03b1\u03b9\u03c9\u03bc\u03ad\u03bd\u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c1\u03b7\u03c4\u03ad\u03c2 \u03b1\u03c0\u03b1\u03b3\u03bf\u03c1\u03b5\u03cd\u03c3\u03b5\u03b9\u03c2 \u03b3\u03b9\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, \u03c0\u03b1\u03c1\u03b1\u03bb\u03ae\u03c0\u03c4\u03b5\u03c2 \u03ba\u03b1\u03b9 destructive actions.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 3<\/span><strong>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03c4\u03bf \u03b5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf evidence<\/strong>\n<p>\u0393\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 UI action \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03c3\u03c4\u03b5 \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03b3\u03b5\u03bd\u03ad\u03c3\u03c4\u03b5\u03c1\u03b7 \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7, \u03c4\u03b1 action parameters, \u03c4\u03bf \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03cc visual clue \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03b1\u03bd\u03b1\u03bc\u03b5\u03bd\u03cc\u03bc\u03b5\u03bd\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1. \u0391\u03c0\u03bf\u03c6\u03cd\u03b3\u03b5\u03c4\u03b5 \u03b1\u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c4\u03b7 \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03c9\u03bd \u03bf\u03b8\u03bf\u03bd\u03ce\u03bd \u03cc\u03c4\u03b1\u03bd \u03b1\u03c1\u03ba\u03b5\u03af redacted evidence.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 4<\/span><strong>\u03a7\u03c9\u03c1\u03af\u03c3\u03c4\u03b5 success \u03b1\u03c0\u03cc safety<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03b5\u03c1\u03ce\u03c4\u03b7\u03c3\u03b7, label \u03ba\u03b1\u03b9 threshold \u03b3\u03b9\u03b1 \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf. \u0388\u03bd\u03b1 task \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b5\u03c4\u03cd\u03c7\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 \u03bd\u03b1 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af block \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b1\u03c0\u03bf\u03ba\u03ac\u03bb\u03c5\u03c8\u03b5 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ae \u03c0\u03b1\u03c1\u03ad\u03ba\u03b1\u03bc\u03c8\u03b5 approval.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 5<\/span><strong>\u03a7\u03c4\u03af\u03c3\u03c4\u03b5 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf reference set<\/strong>\n<p>\u0394\u03ce\u03c3\u03c4\u03b5 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ad\u03c2 trajectories \u03c3\u03b5 \u03b4\u03cd\u03bf reviewers, \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03b4\u03b9\u03b1\u03c6\u03c9\u03bd\u03af\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03c4\u03b5 adjudicated labels \u03b3\u03b9\u03b1 success conditions, \u03c0\u03c1\u03ce\u03c4\u03bf risky step \u03ba\u03b1\u03b9 severity. \u03a7\u03c9\u03c1\u03af\u03c2 reference set \u03b4\u03b5\u03bd \u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03c4\u03b5 \u03c4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 F1 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 domain.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 6<\/span><strong>\u03a4\u03c1\u03ad\u03be\u03c4\u03b5 shadow evaluation<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 precision, recall, false success, missed risk, \u03c7\u03c1\u03cc\u03bd\u03bf \u03b5\u03bd\u03c4\u03bf\u03c0\u03b9\u03c3\u03bc\u03bf\u03cd \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 review \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03c6\u03ae\u03c3\u03b5\u03c4\u03b5 \u03c4\u03bf\u03bd evaluator \u03bd\u03b1 \u03b5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2. \u0391\u03bd\u03b1\u03bb\u03cd\u03c3\u03c4\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac errors \u03b1\u03c0\u03cc perception, consequence inference \u03ba\u03b1\u03b9 aggregation.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 7<\/span><strong>\u0395\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 gates \u03ba\u03b1\u03b9 rollback<\/strong>\n<p>\u0395\u03c0\u03b9\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c7\u03b5\u03b9\u03b1 \u03bc\u03cc\u03bd\u03bf \u03c3\u03c4\u03b9\u03c2 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03ad\u03c2, \u03b1\u03bd\u03b1\u03c3\u03c4\u03c1\u03ad\u03c8\u03b9\u03bc\u03b5\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2. \u0393\u03b9\u03b1 sensitive \u03ae destructive \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03b1\u03c0\u03b1\u03b9\u03c4\u03ae\u03c3\u03c4\u03b5 human handoff, audit trail \u03ba\u03b1\u03b9 rollback, \u03cc\u03c0\u03c9\u03c2 \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 <a href=\"https:\/\/twodots.gr\/terminal-agents-ai-command-line-plires-systima\/\">terminal-agent \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1<\/a>.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<h2 id=\"oria-paragogikis-yiothetisis\">\u03a4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae \u03c5\u03b9\u03bf\u03b8\u03ad\u03c4\u03b7\u03c3\u03b7<\/h2>\n<p>\u039f\u03b9 \u03af\u03b4\u03b9\u03bf\u03b9 \u03bf\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03bf\u03c5\u03bd \u03c4\u03c1\u03b5\u03b9\u03c2 \u03b2\u03b1\u03c3\u03b9\u03ba\u03bf\u03cd\u03c2 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2. \u0397 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03ad\u03c7\u03b5\u03b9 \u03b5\u03c0\u03b9\u03ba\u03c5\u03c1\u03c9\u03b8\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 mobile scenarios\u00b7 \u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03c3\u03b5 desktop \u03ba\u03b1\u03b9 web interfaces \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03ad\u03b1 \u03b5\u03bc\u03c0\u03b5\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7. \u03a4\u03b1 \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 post-hoc, \u03b5\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03bf\u03c5\u03bd \u03b1\u03ba\u03cc\u03bc\u03b7 \u03cc\u03c4\u03b9 \u03c4\u03bf CRATE \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c9\u03c2 online reward model \u03ae real-time reflection module \u03c3\u03b5 \u03ba\u03bb\u03b5\u03b9\u03c3\u03c4\u03cc \u03b2\u03c1\u03cc\u03c7\u03bf.<\/p>\n<p>\u0397 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b4\u03b5\u03bc\u03ad\u03bd\u03b7 \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03b9\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b5\u03c2 \u03c4\u03bf\u03c5 \u03c5\u03c0\u03bf\u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf\u03c5 VLM. \u0391\u03bd \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b4\u03b5\u03bd \u03b1\u03bd\u03b1\u03b3\u03bd\u03c9\u03c1\u03af\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf \u03ae \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c8\u03b5\u03b9 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1, \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03c3\u03c5\u03b3\u03ba\u03ad\u03bd\u03c4\u03c1\u03c9\u03c3\u03b7 \u03b8\u03b1 \u03b2\u03b1\u03c3\u03b9\u03c3\u03c4\u03b5\u03af \u03c3\u03b5 \u03b5\u03bb\u03b1\u03c4\u03c4\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc evidence. \u0395\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd, \u03c3\u03c4\u03bf step-level reasoning \u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03c3\u03c5\u03bd\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c0\u03c1\u03b9\u03bd \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03b4\u03b7\u03bb\u03c9\u03bc\u03ad\u03bd\u03b7 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03ac\u03c1\u03b1 \u03b7 \u03b1\u03be\u03b9\u03bf\u03c0\u03b9\u03c3\u03c4\u03af\u03b1 \u03b1\u03c5\u03c4\u03bf\u03cd \u03c4\u03bf\u03c5 world-model reasoning \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc domain.<\/p>\n<p>\u03a4\u03ad\u03bb\u03bf\u03c2, \u03c4\u03b1 benchmarks \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03c4\u03bf\u03c5 paper \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03c4\u03b9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03ad\u03c2 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd. \u0394\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03bf\u03c5\u03bd \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03bf\u03b9 \u03c4\u03bf\u03c5\u03c2 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03cc workflow. \u0397 \u03c3\u03c9\u03c3\u03c4\u03ae \u03c0\u03bf\u03c1\u03b5\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b9\u03bb\u03bf\u03c4\u03b9\u03ba\u03ae \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03c3\u03b5 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ad\u03c2 trajectories, \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03bc\u03b5 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 labels \u03ba\u03b1\u03b9 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03c4\u03c9\u03bd failure modes \u03c0\u03c1\u03b9\u03bd \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b1\u03c0\u03bf\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9 \u03b5\u03be\u03bf\u03c5\u03c3\u03af\u03b1 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03bf\u03cd\u03c2 \u03bb\u03bf\u03b3\u03b1\u03c1\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd\u03c2 \u03ae \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1.<\/p>\n<h2 id=\"axiologisi-agents-meros-proiontos\">\u0397 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c4\u03c9\u03bd agents \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03af\u03b4\u03b9\u03bf\u03c5 \u03c4\u03bf\u03c5 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2<\/h2>\n<p>\u038c\u03c3\u03bf \u03bf\u03b9 agents \u03bc\u03b5\u03c4\u03b1\u03ba\u03b9\u03bd\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc demos \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2, \u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03ae \u03c4\u03bf\u03c5\u03c2 \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bc\u03b1\u03c4\u03b9\u03ac \u03c3\u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03b7\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03bd\u03ac \u03b2\u03ae\u03bc\u03b1, \u03c3\u03b1\u03c6\u03ae\u03c2 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 task success \u03ba\u03b1\u03b9 safety, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03ba\u03b1\u03b9 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03bf\u03cd\u03bd \u03b1\u03c0\u03cc \u03ac\u03bd\u03b8\u03c1\u03c9\u03c0\u03bf.<\/p>\n<p>\u03a4\u03bf CRATE \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf research blueprint: \u03b1\u03c0\u03bf\u03c3\u03cd\u03bd\u03b8\u03b5\u03c3\u03b7, consequence reasoning \u03ba\u03b1\u03b9 evidence aggregation. \u03a4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 \u03b1\u03c5\u03c4\u03ae \u03b7 \u03b4\u03bf\u03bc\u03ae \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03b9\u03ce\u03c3\u03b5\u03b9 \u03c4\u03bf context overload \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03ba\u03ac\u03bd\u03b5\u03b9 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03b1\u03bd\u03bf\u03b9\u03c7\u03c4\u03cc VLM \u03b1\u03bd\u03c4\u03b1\u03b3\u03c9\u03bd\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03c9\u03c2 evaluator. \u03a4\u03bf CRATE-S \u03b5\u03c0\u03b5\u03ba\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf blueprint \u03c3\u03c4\u03bf\u03bd \u03b5\u03bd\u03c4\u03bf\u03c0\u03b9\u03c3\u03bc\u03cc \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ce\u03bd \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03c9\u03bd, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03c7\u03ad\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03c4\u03b5\u03c5\u03be\u03b7 \u03c3\u03c4\u03cc\u03c7\u03bf\u03c5.<\/p>\n<p>\u0393\u03b9\u03b1 \u03c4\u03bf\u03c5\u03c2 decision makers, \u03c4\u03bf \u03ba\u03b5\u03bd\u03c4\u03c1\u03b9\u03ba\u03cc \u03b5\u03c1\u03ce\u03c4\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc: \u03cc\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1\u03c2 agent \u03bb\u03ad\u03b5\u03b9 \u00ab\u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03ce\u03b8\u03b7\u03ba\u03b5\u00bb, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03b7 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c3\u03bf\u03c5 \u03bd\u03b1 \u03b4\u03b5\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03b1\u03c0\u03b1\u03af\u03c4\u03b7\u03c3\u03b7 \u03b9\u03ba\u03b1\u03bd\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5, \u03bc\u03b5 \u03c0\u03bf\u03b9\u03bf \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf \u03ba\u03b1\u03b9 \u03c3\u03b5 \u03c0\u03bf\u03b9\u03bf \u03b2\u03ae\u03bc\u03b1; \u0391\u03bd \u03cc\u03c7\u03b9, \u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf \u03b2\u03ae\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03ad\u03bd\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf. \u0395\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">\u0391\u03c0\u03cc \u03c4\u03bf trajectory log \u03c3\u03c4\u03bf\u03bd \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae agent<\/p>\n<h3>\u039c\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03c4\u03b1 AI agents \u03c3\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1<\/h3>\n<p>\u0397 TWO DOTS \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03b5\u03b9 success conditions, step-level evidence, safety policies, human approval gates, audit trail \u03ba\u03b1\u03b9 rollback, \u03ce\u03c3\u03c4\u03b5 \u03ba\u03ac\u03b8\u03b5 mobile \u03ae tool-using AI agent \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03c4\u03b9 \u03ad\u03ba\u03b1\u03bd\u03b5 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03bf\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9 \u03b4\u03b9\u03ba\u03b1\u03af\u03c9\u03bc\u03b1 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03c3\u03b5\u03b9.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Check out Business Automation &amp; AI<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf CRATE;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf CRATE \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf VLM-as-judge \u03b4\u03cd\u03bf \u03c3\u03c4\u03b1\u03b4\u03af\u03c9\u03bd \u03b3\u03b9\u03b1 mobile AI agents. \u039c\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 screen-action pair \u03c3\u03b5 \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03bf step-level evidence \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03b1\u03bd \u03cc\u03bb\u03b1 \u03c4\u03b1 \u03c1\u03b7\u03c4\u03ac conditions \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2 \u03b9\u03ba\u03b1\u03bd\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03bf CRATE \u03ba\u03b1\u03b9 \u03c4\u03bf CRATE-S;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf CRATE \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af task completion \u03c3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03b7 \u03c1\u03b7\u03c4\u03ae \u03b5\u03bd\u03c4\u03bf\u03bb\u03ae. \u03a4\u03bf CRATE-S \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af operational safety \u03c7\u03c9\u03c1\u03af\u03c2 task description, \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03c2 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03b9\u03bc\u03bf\u03c5\u03c2 \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5\u03c2 \u03b1\u03bd\u03ac \u03b2\u03ae\u03bc\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae trajectory.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0397 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bf\u03b8\u03cc\u03bd\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03c1\u03cd\u03c8\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c0\u03b1\u03c1\u03ac\u03bb\u03b5\u03b9\u03c8\u03b7, \u03bb\u03ac\u03b8\u03bf\u03c2 \u03c0\u03bf\u03c3\u03cc\u03c4\u03b7\u03c4\u03b1, \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c3\u03b5\u03b9\u03c1\u03ac \u03ae \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf \u03c0\u03bf\u03c5 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c0\u03c1\u03bf\u03c3\u03c9\u03c1\u03b9\u03bd\u03ac. \u0397 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03ba\u03b1\u03b9 \u03b7 \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03b4\u03b9\u03cc\u03c4\u03b7\u03c4\u03b5\u03c2 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7\u03c2 \u03c4\u03b7\u03c2 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a3\u03b5 \u03c0\u03bf\u03b9\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03c4\u03bf AndroidWorld \u03bc\u03b5 116 trajectories, \u03c3\u03c4\u03bf MobileRisk \u03bc\u03b5 204 safe \u03ba\u03b1\u03b9 unsafe trajectories \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf CRATEBench \u03bc\u03b5 187 tasks \u03c3\u03b5 35 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2. \u03a5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf 62 tasks \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c3\u03b5 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03ba\u03b9\u03bd\u03b7\u03c4\u03ae \u03c3\u03c5\u03c3\u03ba\u03b5\u03c5\u03ae.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03b1 \u03ae\u03c4\u03b1\u03bd \u03c4\u03b1 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ac \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039c\u03b5 Qwen2.5-VL, \u03c4\u03bf CRATE \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 F1 0,833 \u03c3\u03c4\u03bf AndroidWorld \u03ba\u03b1\u03b9 0,815 \u03c3\u03c4\u03bf CRATEBench. \u03a4\u03bf CRATE-S \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 F1 0,697 \u03c3\u03c4\u03bf MobileRisk \u03ba\u03b1\u03b9 \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf step-level score \u03b1\u03c0\u03cc \u03c4\u03bf OS-Sentinel \u03c3\u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c1\u03cd\u03b8\u03bc\u03b9\u03c3\u03b7.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03b5\u03af \u03c3\u03ae\u03bc\u03b5\u03c1\u03b1 \u03c3\u03b5 web \u03ae desktop agents;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0397 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03b1\u03b8\u03bf\u03b4\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03bf evaluator, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03c0\u03c1\u03bf\u03b4\u03b7\u03bc\u03bf\u03c3\u03af\u03b5\u03c5\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03b5\u03c0\u03b9\u03ba\u03c5\u03c1\u03ce\u03c3\u03b5\u03b9 \u03c4\u03bf CRATE \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 mobile scenarios \u03ba\u03b1\u03b9 post-hoc \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7. Web, desktop \u03ba\u03b1\u03b9 real-time \u03c7\u03c1\u03ae\u03c3\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bd\u03ad\u03b1 \u03b5\u03bc\u03c0\u03b5\u03b9\u03c1\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03b2\u03b5\u03b2\u03b1\u03af\u03c9\u03c3\u03b7.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03c5\u03bd\u03b4\u03ad\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1\u03c2 \u03ae \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5 \u03bc\u03b5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 evidence. \u0391\u03c5\u03c4\u03cc \u03b2\u03bf\u03b7\u03b8\u03ac \u03c4\u03bf auditing, \u03c4\u03bf failure analysis, \u03c4\u03b1 approval gates \u03ba\u03b1\u03b9 \u03c4\u03bf human handoff, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5 \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03b1\u03c3\u03c6\u03b1\u03bb\u03bf\u03cd\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf\u03c2 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03cc\u03c2 \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf\u03c2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f evaluator \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03cd\u03c3\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03bf\u03b8\u03cc\u03bd\u03b7 \u03ae \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03bc\u03b9\u03b1\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03ba\u03c1\u03af\u03c3\u03b7 \u03bd\u03b1 \u03ba\u03bb\u03b7\u03c1\u03bf\u03bd\u03bf\u03bc\u03ae\u03c3\u03b5\u03b9 \u03c4\u03bf \u03bb\u03ac\u03b8\u03bf\u03c2. \u0391\u03c0\u03b1\u03b9\u03c4\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 domain-specific labels, privacy-safe logs, shadow tests \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b5\u03c7\u03ae\u03c2 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 false positives \u03ba\u03b1\u03b9 false negatives.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.20797\" target=\"_blank\" rel=\"noopener\">arXiv \u2014 CRATE: Automated Trajectory Evaluation for Mobile Agents<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2510.24411\" target=\"_blank\" rel=\"noopener\">arXiv \u2014 OS-Sentinel \u03ba\u03b1\u03b9 MobileRisk \u03b3\u03b9\u03b1 operational safety<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/google-research\/android_world\" target=\"_blank\" rel=\"noopener\">Google Research \u2014 \u03b5\u03c0\u03af\u03c3\u03b7\u03bc\u03bf repository \u03c4\u03bf\u03c5 AndroidWorld<\/a><\/li>\n<li><a href=\"https:\/\/www.nist.gov\/itl\/ai-risk-management-framework\" target=\"_blank\" rel=\"noopener\">NIST \u2014 AI Risk Management Framework \u03b3\u03b9\u03b1 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03ba\u03c5\u03b2\u03ad\u03c1\u03bd\u03b7\u03c3\u03b7 \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf CRATE \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03ad\u03bd\u03b1\u03bd mobile AI agent \u03b1\u03bd\u03ac \u03b2\u03ae\u03bc\u03b1, \u03be\u03b5\u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03bf\u03c1\u03b3\u03b1\u03bd\u03ce\u03bd\u03b5\u03b9 evidence, audit \u03ba\u03b1\u03b9 approval gates.<\/p>","protected":false},"author":1,"featured_media":95392,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[6506,18337,20017,18467,20018],"class_list":["post-92253","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-agents","tag-ai-safety","tag-mobile-agents","tag-multimodal-ai","tag-vlm-evaluation"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/92253","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=92253"}],"version-history":[{"count":1,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/92253\/revisions"}],"predecessor-version":[{"id":95396,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/92253\/revisions\/95396"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/95392"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=92253"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=92253"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=92253"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}