Meta's GitSwarm: Hundreds of Async Agents Sharing a Git Repo Solve All 30 IMOProofBench-Advanced Problems
omarsar0 · x · 2026-10-11
A new paper from Meta Superintelligence Labs tackles compounding inference for long-horizon agents. Key insight: inference-time compute usually goes to separate attempts that discard partial work and failed ideas. GitSwarm gives hundreds of asynchronous agents one shared Git repository — each contribution is a commit recording which prior commits it builds on, with no central planner; each worker decides whether to explore, verify, refine, or combine. With GPT-5.5 Codex workers it solves all 30 IMOProofBench-Advanced problems in one run and scores 79.4% on ProgramBench vs 65.1% for a single persistence-prompted Codex agent. Persistent shared memory lets compute actually compound.
More from coding & agent
- GPT Researcher crosses 30k GitHub stars, built and maintained almost entirely by community — EdenEmarco177 · 2026-10-11
- shadcn: AI hasn't sped up abstraction design yet, it just helps me fail faster — shadcn · 2026-10-11
- Free 7-week production Agentic RAG course hits 9.7k stars on GitHub — mdancho84 · 2026-10-11
- Power User Runs AI Coding Agent 24/7, Burns Through Weekly Limit in Two Days — gandamu_ml · 2026-10-11
- Hallucinated model name sat in codebase for months, caught 6 days before a deprecated model shutdown — Trout_dev · 2026-10-11
- Build a multiplayer coding agent on Slack in 4 steps with eve — cramforce · 2026-10-11