GitSwarm: Decentralized Compounding Inference via Shared Git Repositories

Anirudh Goyal and colleagues (authors include Salakhutdinov, Courville, and Jason Weston) released the paper "GitSwarm: Decentralized Compounding Inference," proposing a "compounding inference" framework: instead of discarding each agent run as a one-off trajectory, it lets runs accumulate as persistent work assets that other agents can inspect, extend, combine, or challenge, with a Git repository as the shared substrate. What makes the paper notable: it tackles a fundamental dilemma in inference-time scaling and uses real research experiments to demonstrate spontaneous collaboration and cross-time knowledge reuse in decentralized agent swarms.

Confirmed

• The authors point out the dilemma in current inference-time scaling: running more independent attempts leads to work being rediscovered over and over; stretching a single agent trajectory leads to context compression and constant loss of discoveries. Neither treats intermediate research results as first-class citizens.
• GitSwarm is decentralized by design: no central planner, no fixed roles, no privileged final arbiter; all workers are fully homogeneous, make autonomous decisions, and run asynchronously, coordinated only by a lightweight scheduler that maintains concurrency — drawing a clear line from traditional planner/role-based multi-agent systems.
• Git's deeper value is redefined: branches preserve competing hypotheses; commits make experiments and failures inspectable; explicit informedby links record semantic inheritance when ideas are borrowed across branches.
• In open-ended AutoResearch experiments, agent swarms autonomously proposed modifications to three starting models: Residual Matrix Transformers, Looped Transformers, and NanoChat.
• The team observed collaboration patterns emerging without assigned roles: workers continuously built and integrated on ProgramBench coding tasks; on proof tasks, effort spontaneously shifted toward verification and candidate filtering — evidence of the decentralized design's adaptability.
• A representative case: in the RMT experiment, an agent revisited an experiment that had failed 49.5 hours earlier, and the negative result inspired the design of a "multi-token memory mixer" that was ultimately implemented.

Why it matters

• The work turns agent run logs into branchable, attributable persistent assets that record what worked, what failed, and what to try next, offering a mechanistic solution for long-horizon autonomous research and cross-agent knowledge reuse.

2026-10-10 ~ 2026-10-10 · 10 related posts

• [source] GitSwarm proposes "compounding inference": treating agent runs as reusable Git-based artifacts — anirudhg9119 · 2026-10-10
• The inference scaling dilemma: independent attempts rediscover work, long trajectories lose discoveries — anirudhg9119 · 2026-10-10
• How GitSwarm works: agents take over a shared Git repo, inheriting a growing research history — anirudhg9119 · 2026-10-10
• Git is more than storage: branches preserve hypotheses and commits make failures inspectable in GitSwarm — anirudhg9119 · 2026-10-10
• No planner, no roles, no judge: GitSwarm runs identical async workers with just a lightweight scheduler — anirudhg9119 · 2026-10-10
• GitSwarm's open-ended AutoResearch: agent swarms autonomously improve transformer architectures on GPUs — anirudhg9119 · 2026-10-10
• A failed experiment from 49.5 hours earlier inspired a multi-token memory mixer in GitSwarm's final model — anirudhg9119 · 2026-10-10
• GitSwarm agents spontaneously adapt their workflow: building on coding tasks, verifying on proof tasks — anirudhg9119 · 2026-10-10
• [source] Compounding inference paper with Munos, Fergus, Weston, Courville: branchable agent reasoning records — anirudhg9119 · 2026-10-10
• [source] GitSwarm paper: agents collaborate through a shared Git repo, solving all 30 IMOProofBench-Advanced problems — anirudhg9119 · 2026-10-10