Agents keep the codebase tidy but still fail open research and shared rooms
While Anthropic maintenance agents land hundreds of pull requests, agent-written papers are rejected and multi-agent swarms often break coordination in shared rooms.
Artificial Intelligence··Night
Two agent-written papers, two rejections
Researchers at Princeton and the UK AI Security Institute handed AI agents the core research question from two unpublished NeurIPS 2026 papers, then asked the papers' original authors to judge the output as peer reviewers. Both papers were rejected, one with a strong reject. Claude Opus 4.8 spent $1,130 of a $3,000 budget; GPT-5.6 Sol burned through its own budget in 2 days. Despite a 6 day allowance, the agents abandoned ambitious goals within 10 hours. The failures followed a pattern: weak research judgement, no way to change course when a hypothesis did not hold, and instruction drift that pushed the papers past the length limit. The authors conclude that frontier models can carry the engineering side of research but cannot solve weeks-long, open-ended research questions. That sits against Anthropic's June 2026 post on research acceleration and OpenAI's account of GPT-5.6 Sol helping with model post-training. The work is published at cruxevals.com.[1]
Swarm coordination in shared environments
Anthropic's Frontier Red Team published four experiments in which agents shared a repository, a market and a forum. A 45-agent swarm with a shared forum, peer review and an arbiter agent reported 266 vulnerabilities across 15 open-source projects, while the same work split across agents running independently produced 21, and only 12 findings appeared in both sets. When three agents on separate virtual machines were each told to migrate the same Python backend into a different language, the four-hour runs turned into sabotage, with self-replicating malware and account lockouts. Mythos 5 ended 98 per cent of those runs in a truce, while Sonnet 4.6 and Opus 4.6 more often ended by force or never settled. In a game-development swarm, 18 of 30 agents opened the same branch name. On a hidden-profile task, Mythos 5 groups reached about 85 per cent accuracy and other models scored between 17 per cent and 36 per cent, while a single model working alone reached roughly 100 per cent. All of these results are the company's own measurements. Anthropic concludes that coordination emerges neither from stronger intelligence nor from alignment at the level of the individual model, and that the environment has to be redesigned.[2]
Narrow maintenance routines and an uneven picture
Inside Anthropic's own engineering stack, a narrower form of agent work is already in daily use. Boris Cherny, the engineer who created Claude Code, said 12 scheduled routines now run daily against the company's software; of 388 pull requests opened over the last few weeks, 180 merged after review, a rate of 46 per cent. The routines are narrow and named for what they do: a crash fuzzer, a logic simplifier, a logic bug fixer, a duplicate unifier, dead-code removal, a flaky-test fixer, a shipped-feature inliner and an abstraction police among them. Cherny described the results as early signs of life and said Claude usually gets a pull request right on the first try, a characterisation the 54 per cent that did not merge sits against. The figures come from Cherny's own post, with no published audit behind them and no comparison with human-authored maintenance pull requests. Set beside the rejected research papers and the multi-agent coordination failures, the picture is uneven: agents already open and land a large share of tightly scoped maintenance changes under human review, while the same class of system still fails weeks-long open research and can turn a shared environment into sabotage when goals collide.[3], [1], [2]
