[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-ai-agents-that-teach-themselves-skills-can-get-worse":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},6937,"ai-agents-that-teach-themselves-skills-can-get-worse","AI Agents That Teach Themselves Skills Can Get Worse","New research shows AI agents that write their own skills eventually contaminate their knowledge base, and filtering skills before they're added can stop it.","Letting an AI agent write its own playbook sounds efficient, until the playbook starts poisoning itself.\n\nA new arXiv paper looks at self-evolving agents that build up a library of reusable \"skills\" distilled from their own past task runs. The researchers found this accumulation isn't a straight line up. Past a certain pool size, adding more skills makes the agent perform worse, not better. The cause is a defective skill entering the agent's decision context and then becoming the template other skills get distilled from, spreading the flaw across rounds. Worse, deleting the bad skill afterward barely helps, because its descendants already inherited the broken reasoning.\n\nThat irreversibility is the real finding here. It means skill quality has to be checked before a skill ever reaches the runtime library, not patched after the fact. The paper's fix, called Verifier-as-Gatekeeper, runs three separate checks - structural validity, behavioral harmlessness, and semantic consistency - plus a step that trims redundant skills at the top of the pool. On the Terminal-Bench 2 benchmark, this pushed accuracy to 72% pass@1 using a skill pool roughly five times smaller than the unfiltered version, and the resulting pool worked on four other AI backbones without retraining.\n\nIt's a useful reminder that \"let the agent improve itself\" is not a free lunch: self-evolution needs the same input validation any codebase does, or it just compounds its own mistakes faster.","[\"ai-agents\",\"llm\",\"ai-safety\",\"research\"]","2026-09-18T04:00:00.000Z","2026-09-18T23:31:55.763Z","2026-09-18T23:32:07.685Z","published",null,[],"ai",[26,27,28,29],"ai-agents","llm","ai-safety","research",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.05810",0,{"sections":36},[37,40,44,49,54,58,62,67,71,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",4082,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",661,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",339,"2026-09-17T12:00:00.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":18},"Hardware","hardware",155,{"name":59,"slug":60,"count":61,"latest_published_at":18},"Science","science",125,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":18},"Dev Tools","dev-tools",78,{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",55,"2026-09-09T23:14:29.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",43,"2026-09-10T12:18:06.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",41,"2026-09-08T01:57:23.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]