[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-optimizer-lets-ai-agents-rewrite-their-own-prompts-and-tools":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},9884,"new-optimizer-lets-ai-agents-rewrite-their-own-prompts-and-tools","New Optimizer Lets AI Agents Rewrite Their Own Prompts and Tools","VERSE, a new optimizer, improves AI coding agents by testing and verifying its own edits before applying them, beating prior methods on benchmark tasks.","Researchers built an optimizer that doesn't just rewrite an AI agent's instructions - it rewrites its own toolkit for figuring out what's broken.\n\nThe project, called VERSE, tackles a gap in \"harness evolution\" - the practice of automatically improving an agent's prompts, tools, and workflow without retraining the model - where the optimizer doing the tuning normally keeps its own methods fixed. The researchers found that letting the optimizer evolve itself backfired unless it could verify each change by actually running it; unverified self-evolution made results worse, not better. VERSE builds that verification in: before submitting an edit, it tests the draft, replays past failures, and perturbs suspected steps, tracking which rounds fix problems and which create new ones. It applies this process to both the agent's harness and its own prompts, skills, tools, hooks, and notes, while the weights of both models stay untouched.\n\nThat distinction matters because harness tuning, not model retraining, is increasingly how teams squeeze more performance out of agents they can't afford to fine-tune. VERSE suggests that tuning process can itself be automated and checked, rather than left to a static optimizer script. On held-out SWE-rebench coding tasks and newer out-of-distribution tasks across five languages, VERSE-based optimizers beat the strongest baselines - 42.3% versus 39.2% accuracy on one benchmark, 37.7% versus 29.3% on the other.\n\nThose are still double-digit swings on benchmark tasks, not production codebases - a reminder that \"self-evolving\" optimizers are, for now, mostly getting better at a narrow kind of homework.","[\"ai agents\",\"agent harnesses\",\"benchmarks\",\"ai research\"]","2026-10-05T04:00:00.000Z","2026-10-05T11:52:08.693Z","2026-10-05T11:52:13.939Z","published",null,[],"ai",[26,27,28,29],"ai agents","agent harnesses","benchmarks","ai research",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02616",0,{"sections":36},[37,40,44,49,54,59,63,68,73,78,83,88,93,98],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",6154,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",859,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",444,"2026-10-03T15:02:01.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",323,"2026-10-04T13:00:00.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",204,"2026-10-03T14:50:50.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":18},"Science","science",177,{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",158,"2026-10-03T03:21:12.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",97,"2026-10-04T10:00:00.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":77},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",92,"2026-10-04T14:36:25.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",51,"2026-10-05T02:35:01.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",32,"2026-10-02T18:00:00.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]