[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-framework-pinpoints-which-ai-agent-step-caused-a-failure":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},9517,"new-framework-pinpoints-which-ai-agent-step-caused-a-failure","New Framework Pinpoints Which AI Agent Step Caused a Failure","DeFA maps cause and effect across an AI agent's actions to pinpoint which step and agent actually caused a failure, not just where it surfaced.","Researchers have built a system that traces exactly which step in a long AI agent task went wrong, and why, instead of just confirming that something did.\n\nThe tool, called DeFA, builds a dependency graph of every action an agent takes, linking each step to the ones before and after it both by protocol (what the task required) and by meaning (what the agent was trying to do). When a step violates the task's requirements, DeFA traces that violation back to its source and forward to its consequences, producing a \"failure propagation graph\" that shows how one bad step cascades into a visible error much later. For agent runs too long to analyze in one pass, DeFA splits the trajectory into segments, pairing detailed analysis of the current segment with summaries of the rest so it never loses the global picture. On the Who and When benchmark and its harder Pro subset, DeFA came out most accurate at naming both the responsible agent and the exact failing step, across every model backbone tested.\n\nThat precision matters because most agent-failure analysis today stops at \"the agent got the wrong answer,\" leaving developers to manually replay transcripts to find the actual breakpoint. DeFA's diagnoses were also fed into a skill-improvement pipeline called Trace2Skill, where they lifted downstream task accuracy by 6 to 15 percentage points over that pipeline's native error handling - turning root-cause analysis into a training signal, not just a debugging aid.\n\nMulti-step AI agents are multiplying faster than anyone's ability to explain why they fail, and unglamorous plumbing like this - plain cause-and-effect attribution instead of a single pass or fail score - is what will actually decide whether anyone trusts them in production.","[\"ai agents\",\"llm\",\"failure attribution\",\"research\"]","2026-10-02T04:00:00.000Z","2026-10-02T22:50:00.615Z","2026-10-02T22:50:05.390Z","published",null,[],"ai",[26,27,28,29],"ai agents","llm","failure attribution","research",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.01256",0,{"sections":36},[37,40,44,48,53,57,61,66,71,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",5886,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",835,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",438,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":18},"Hardware","hardware",199,{"name":58,"slug":59,"count":60,"latest_published_at":18},"Science","science",171,{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]