arXivAnalysisResearchers release GitSwarm framework for compounding multi-agent inference
What happened.Researchers introduced GitSwarm, an open framework that organizes autonomous reasoning agents around a shared, branchable Git repository. The system lets agents save partial solutions and failed attempts so subsequent workers can build on earlier progress.
Why it matters.Traditional agent loops discard failed reasoning runs, driving up token bills on long research tasks. Founders running coding swarms can cut inference spend by caching intermediate branch state in git trees instead of restarting prompts from scratch.
The technical specifications:
- Architecture: Autonomous workers record intermediate code and proofs as atomic commits with explicit semantic dependencies.
- Performance: Solved all 30 problems on IMOProofBench-Advanced in a single run using GPT-5.5 as the underlying reasoning model.
- Reusability: Achieved 79.4% on ProgramBench with a 94.7% reuse rate for intermediate contributions across parallel branches.