--- title: 'In the News: November 9, 2026 (Evening)' description: 'Anthropic says 1,000 agents coordinated through Git and a shared SQL database. A harness benchmark finds pass rates tie while costs differ.' canonical_url: 'https://darkfactory.dev/news/2026-11-09-evening' markdown_url: 'https://darkfactory.dev/news/2026-11-09-evening.md' collection: news date_published: '2026-11-09T19:05:00-05:00' date_modified: '2026-11-09T19:05:00-05:00' --- # In the News: November 9, 2026 (Evening) Anthropic published the full account of its 1,000-agent Fermat's Last Theorem run, and the useful part is the coordination design: no messaging tools, only Git branches and a SQL database the lead agent set up in ten minutes. Separately, an open harness benchmark finds that four agent harnesses pass about the same share of tasks on Opus 5.5 while their cost per solved task differs by a factor of two. ## 1. Anthropic details how 1,000 agents coordinated through Git and Postgres **[Safely scaling multi-agent teams with Claude Fable 5.2](https://claude.dev/blog/safely-scaling-multi-agent-teams-with-fable-5-2/)** · Barry Zhang, Tianyi Peng, Elle Najt, Keller Scholl and Lance Martin, Anthropic · claude.dev, November 9, 2026 Anthropic describes what it calls a toy experiment in which 1,000 agents running an early version of Fable 5.2 reproduced a formal proof of Fermat's Last Theorem. The team took 28.7 hours, used one-third as many output tokens and wrote a proof one-seventh the length of the September formalization, which Anthropic says took a researcher and dozens of agents 11 days and about 6 billion output tokens. This time there was no Prove2Me platform and no researcher steering. Each agent had its own sandbox with no network, a shell, Lean and Mathlib, a Git branch with merge requests into main, and a shared Postgres database. The prompt named one lead and 999 helpers and gave a list of 60 statements to prove. The lead built the coordination layer itself. In the first ten minutes it created five tables: announce, assign, notes, subtasks and landed. Helpers claimed tasks from the subtasks table and logged finished work. Early on the helpers judged 47 of the 60 statements infeasible within the run's budget, and the lead set those aside; the team went on to prove most of them. For the last and largest statement, Langlands-Tunnell, a captain had put the chance of finishing below 1%, and the lead wrote "Honest odds are low." as it moved about 750 helpers onto it. The helpers grew a task tree of 2,418 tasks, up to 10 levels deep. Anthropic also says a single shared machine hit its limits at about 100 agents, and more than half of its 100-agent runs had to be stopped or restarted. On safety, a classifier checked every tool call before it ran and an offline monitor reviewed the team's actions. A block on the lead would have ended the run; a block on a helper removed only that helper. Across more than a million tool calls, none met the bar to be blocked. Anthropic adds that it runs such experiments on models that may not have finished its pre-release alignment assessment. **Why it matters:** If you run parallel agents, this is a concrete coordination pattern you can copy: separate sandboxes, Git for code, and plain SQL tables for claims, notes and completed work, which Anthropic chose because transactions handle concurrent writes. The post also shows agents misjudging difficulty early, and Anthropic says human encouragement at such moments can still keep agents from stopping too early. Every figure is Anthropic's own from one described experiment, and we read the post, not the underlying logs or the September formalization. ## 2. A new benchmark holds the model fixed and swaps the harness **[harness-evals](https://github.com/garrytan/g-harness-evals)** · Garry Tan, Y Combinator · GitHub, repository README read November 9, 2026 The repository holds Opus 5.5 and a set of 20 public multi-user tasks fixed and changes only the harness. On Opus 5.5, pass rates were QM 72.5%, Claude Code 76.3%, Hermes Agent 78.8% and OpenClaw 76.3%, and no tested pair could be told apart. Cost per solved task was $0.19 for QM, $0.22 for Claude Code, $0.24 for Hermes Agent and $0.44 for OpenClaw. The README says about 81% of OpenClaw's extra cost over Claude Code comes from cache writes. The per-harness findings are more specific than the ranking. In Claude Code, a rule placed in the system prompt or startup file was followed in 30 of 30 turns against 18 of 30 with no rule, and a skill that was not mentioned was opened 0 times in 30 turns. Dropping its memory-and-skills text together with its tool-description advice took pass rate on harder tasks from 60.4% to 0.0%. Told to run gstack's skills, Opus 5.5 loaded the skill text every time and ran none of the steps, in 0 of 46 sessions. **Why it matters:** The cost gap is large enough to matter at volume, and the placement result is cheap to act on: put the rules you need followed where the agent sees them every turn, and name the skills you want used. The README labels these dev results on public tasks, not held-out confirmation, and says most differences are descriptive. Tan runs Y Combinator, which the README says owns QM and invests in Hermes, and gstack and gbrain are his own projects. We did not open the linked per-finding reports or rerun the tasks. --- ## Also this cycle - **[CodeRabbit scales agentic software development and ships 20x more](https://openai.com/index/coderabbit-codex-daybreak/)** · OpenAI · An OpenAI-published interview with CodeRabbit CEO Harjot Gill, who says the company has shipped roughly 20 times more software over six months with the same team using Codex. He says the first bottleneck after adopting coding agents is code review, and that the team reworked its release process and where human approvals sit, auto-shipping low-risk changes. These are the company's own estimates, and CodeRabbit sells code review tooling.