Meta's GitSwarm: Git repo as shared memory for agent swarms, solving IMO-ProofBench in one run
armand_ruiz · x · 2026-10-13
Meta Superintelligence Labs, with Mila, CMU and Princeton, released GitSwarm: Decentralized Compounding Inference, an async multi-agent system that makes inference-time compute compound instead of restarting each run:
- No central planner: homogeneous async agents independently decide to explore, verify, refine, or combine prior work
- Git repo as shared memory: every contribution is an atomic commit, branchable, with semantic dependencies recording lineage
- ProgramBench: 79.4% mean score vs 65.1% for the strongest baseline; scaling from 73.1% at 100 workers to 79.4% at 500
- IMOProofBench-Advanced: all 30 problems solved in one run with GPT-5.5, under 1M output tokens
- Reuse is real: 94.7% of contributions were later built upon; the winning solution's ancestry covers 82–93% of the contribution graph
Takeaway: inference-time computation can accumulate across independent episodes into an evolving, reusable body of work.
More from coding & agent
- Enterprise AI's Worst Failure Mode: Confident Wrong Answers, No Errors — DavidLinthicum · 2026-10-13
- Dev builds Felix, a detective game living on an M5Stack device with Codex and receipt printer clues — OpenAIDevs · 2026-10-13
- Muse Wonder tops SWE-Marathon at 57.5%, beating Claude and GPT-6 by 11+ points — jyangballin · 2026-10-13
- 72% of ReadMe's docs traffic is agents; Restless puts API fixes inside error responses — JaynitMakwana · 2026-10-13
- Mochi: an open-source desktop companion that watches your Claude Code agents for you — brute984 · 2026-10-13
- Perplexity built Projects' filesystem as a Git forge for agents, with greps beating semantic search — perplexity_ai · 2026-10-13