Meta's GitSwarm: agents collaborating via shared Git repo hit 79.4% vs 65.1% baseline
rohanpaul_ai · x · 2026-10-11
A Meta-led paper introduces compounding inference: organizing inference-time compute so intermediate work — including failures — persists and can be extended by later computation.
- GitSwarm: asynchronous swarm of homogeneous agents collaborating through a shared, branch-able Git repo with no central task assigner; atomic commits preserve artifacts, and explicit semantic dependencies record what each contribution builds on.
- Results: 79.4% mean on ProgramBench's 50 program-rebuilding tasks vs 65.1% for the strongest single-agent baseline at similar compute; later agents reused 94.7% of saved contributions. On IMOProofBench-Advanced it solved all 30 problems in one run with GPT-5.5.
- Takeaway: instead of pushing one agent to keep going, run many over a shared Git history that keeps every attempt, failures included.
More from coding & agent
- Michael Black: agents delivering results daily feels like Christmas morning — Michael_J_Black · 2026-10-11
- Heavy coding-agent user on decision fatigue: headaches fade with adaptation — dejavucoder · 2026-10-11
- Arena scores look close: OpenAI 88 vs Claude 83 means double the error rate — i_dg23 · 2026-10-11
- Token prices keep falling, yet devs burn more: Jevons paradox hits AI coding agents — daniel_mac8 · 2026-10-11
- Qwen Code Desktop v0.25.1 ships session diagnostics fix and MCP image handling improvements — github-actions[bot] · 2026-10-11
- vboykis: you can't get good AI outcomes without years of writing code first — vboykis · 2026-10-11