[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-self-editing-ai-agents-fail-to-undo-most-of-their-own-changes":10,"sections":46},{"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":35,"tags":36,"sources":41,"feedback":45,"feedback_at":22,"cost_usd":45,"total_tokens":45},6674,"self-editing-ai-agents-fail-to-undo-most-of-their-own-changes","Self-Editing AI Agents Fail to Undo Most of Their Own Changes","A new benchmark finds AI agents that edit their own tools usually cannot undo the changes without precise state tracking and a richer recovery vocabulary.","Researchers built a stress test for AI agents that rewrite their own prompts and tools - and most of the rewrites cannot be undone.\n\nThe paper, called EvoUndo, ran 600 one-shot self-evolution tasks in which an LLM agent modified its own prompts, tools, or middleware. Of those, 197 mutations improved the agent's capability but failed a recoverability check, meaning the agent could not cleanly reverse them once the system had moved to a different state. Standard repair methods fixed none of the 197. A deterministic oracle then split those failures into two groups: 48 were fixable using the agent's original recovery language, and 143 needed an expanded recovery vocabulary to be fixable at all - together covering 191 of the 197 cases, with 6 unrecoverable by any method tested.\n\nWhat closes each gap is the more interesting finding. Giving the agent exact state-address tracking - knowing precisely which resource changed and where - lifted practical recovery in the 48-case group from 0 to 38, a 79 percent success rate. Swapping in the richer recovery language fixed 142 of the 143 harder cases. But stacking both fixes together on one model backbone, gpt-oss-120b, actually dropped that number to 133, a regression that disappeared when the same setup ran on a different model, Qwen3.8-27B.\n\nLetting an agent patch itself has never been the hard part. Making sure it can also cleanly un-patch itself, on demand, in a system that has since moved on - that is the part nobody had actually tested until now.","[\"ai-agents\",\"ai-safety\",\"llm-research\",\"dev-tools\"]","2026-09-17T04:00:00.000Z","2026-09-18T05:32:29.736Z","2026-09-18T05:32:41.650Z","published",null,[24,30],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Add the paper's citation (arXiv:2608.28363) so this benchmark study isn't attributed only to unnamed 'researchers' with no link for readers to verify.","resolved",{"id":31,"reviewer":32,"round":33,"reason":34,"status":29},"publisher-r2","publisher",2,"The ablation numbers don't reconcile with the headline stat — the piece never explains how the 48-case and 143-case subsets relate to the 197 total, making the reported recovery rates internally inconsistent.","ai",[37,38,39,40],"ai-agents","ai-safety","llm-research","dev-tools",[42],{"name":43,"url":44},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.28363",0,{"sections":47},[48,52,56,61,66,70,74,79,84,87,92,97,102,107],{"name":49,"slug":35,"count":50,"latest_published_at":51},"AI",3853,"2026-09-17T08:27:09.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":18},"Security","security",648,{"name":57,"slug":58,"count":59,"latest_published_at":60},"Policy","policy",338,"2026-09-11T04:00:00.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":18},"Hardware","hardware",154,{"name":71,"slug":72,"count":73,"latest_published_at":18},"Science","science",114,{"name":75,"slug":76,"count":77,"latest_published_at":78},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":85,"slug":40,"count":86,"latest_published_at":18},"Dev Tools",73,{"name":88,"slug":89,"count":90,"latest_published_at":91},"Startups","startups",55,"2026-09-09T23:14:29.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Gaming","gaming",43,"2026-09-10T12:18:06.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"General","general",41,"2026-09-08T01:57:23.000Z",{"name":103,"slug":104,"count":105,"latest_published_at":106},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":108,"slug":109,"count":110,"latest_published_at":111},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]