[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-13-ai-agents-shared-a-git-history-to-rebuild-a-model":10,"sections":45},{"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":40,"feedback":44,"feedback_at":22,"cost_usd":44,"total_tokens":44},6597,"13-ai-agents-shared-a-git-history-to-rebuild-a-model","13 AI Agents Shared a Git History to Rebuild a Model","Thirteen agentic AI workers shared a Git-based memory to close 62% of a frozen model's performance gap, but the paper's contributor count doesn't add up.","Thirteen AI agents with no boss and no assigned tasks spent almost 12 days rebuilding a broken language model by leaving notes for each other in Git.\n\nA new system called Agora turns Git into shared memory for autonomous research agents: every hypothesis, experiment, and verification becomes an immutable commit, with parent links showing what each result builds on and an index tracking neglected branches. In its first extended test, 13 language-model agents ran for nearly 12 days on one problem: initialize a frozen 119.6-million-parameter model built on an architecture that matched none of 141 available donor models, without training data or gradient updates. The agents filed 1,703 contributions and pushed the model's error score from 3.39 down to 1.899 bits per byte, closing 62% of the gap to a fully trained GPT-2 124M. The winning method copied statistical patterns from donor models into the target's embedding and output layers, then added short-range context through targeted edits to its attention, feed-forward, and state-space components.\n\nThis is a real test of a specific idea: agents that read each other's failed and successful attempts should out-research agents working in isolation, which is how most autonomous research loops operate today. The paper reports 165 independent attempts to reproduce the winning recipe, all of which succeeded - a rare clean sweep in AI research, let alone agent-generated research. One number doesn't hold up under scrutiny: the winning recipe's 145-commit history is credited to 15 separate accounts, two more than the 13 agents that did the work, and the paper never explains where the extra accounts came from.\n\nThe authors themselves say this run shows consensus more than it proves collaboration helps; the controlled comparison against solo agents working the same problem, the one that would show whether a shared Git log beats working alone, still hasn't been done.","[\"ai\",\"ai-agents\",\"machine-learning\",\"research\"]","2026-09-17T04:00:00.000Z","2026-09-18T01:54:56.357Z","2026-09-18T01:55:08.280Z","published",null,[24,30],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"The claim that '165 people reran the winning recipe independently' misattributes the source's '165 independent reproductions' to human reproducers in an experiment that is otherwise entirely about autonomous AI agents with no human involvement — clarify whether these reproductions were run by other agents\u002Fautomated instances or actual humans before publishing.","resolved",{"id":31,"reviewer":32,"round":33,"reason":34,"status":29},"publisher-r2","publisher",2,"The article says 13 agents produced the work but attributes the winning recipe's 145 commits to 15 accounts without explaining the discrepancy, an unresolved numeric inconsistency.","ai",[35,37,38,39],"ai-agents","machine-learning","research",[41],{"name":42,"url":43},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.18094",0,{"sections":46},[47,51,55,60,65,69,73,78,83,87,92,97,102,107],{"name":48,"slug":35,"count":49,"latest_published_at":50},"AI",3853,"2026-09-17T08:27:09.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":18},"Security","security",648,{"name":56,"slug":57,"count":58,"latest_published_at":59},"Policy","policy",338,"2026-09-11T04:00:00.000Z",{"name":61,"slug":62,"count":63,"latest_published_at":64},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":18},"Hardware","hardware",154,{"name":70,"slug":71,"count":72,"latest_published_at":18},"Science","science",114,{"name":74,"slug":75,"count":76,"latest_published_at":77},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":18},"Dev Tools","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"]