GitSwarm: Decentralized Compounding Inference

An agent may spend minutes reconstructing a codebase, testing a hypothesis or finding why an implementation fails. If its context ends and the result is not preserved in a structured form, the next agent repeats the same exploration. A shared chat transcript can retain text but often mixes hypotheses, evidence, code, decisions and stale context. A summary compresses information but can erase alternatives and provenance. GitSwarm treats the repository as structured memory: artifacts remain inspectable, competing lines of work remain separate, and dependencies are recorded so future workers can discover how a result was obtained.

Role-based systems such as MetaGPT and AutoGen commonly structure collaboration through assigned roles, message passing or an orchestrator. Independent best-of-N sampling explores candidates but typically discards non-selected work. Long single-agent trajectories preserve work inside one context but face context limits and can lose structured provenance. Summary-based memory compresses prior work but may flatten alternatives and omit exact artifacts. GitSwarm's differentiator is a persistent branchable artifact store with explicit cross-branch semantic dependencies and worker-selected next actions. It complements orchestration frameworks rather than making them obsolete: a production system could use an orchestrator for policy or routing while adopting Git-style durable memory for work products.

The system has three principal parts. A pool of homogeneous workers runs the same model and behavioral contract. A shared, branchable Git repository stores the task, code, experiments, findings and candidate solutions. An execution harness manages worker scheduling, isolated workspaces, publication validation, live activity visibility and final selection. The harness coordinates execution but does not assign intellectual roles such as researcher, tester or integrator. Each worker chooses what to do after inspecting the evolving repository. Specialization can emerge from the state of the work rather than from a centrally assigned role taxonomy.

The strongest research-aligned keyword cluster is around agent harnesses and persistent collaboration. Prioritise: compounding inference, multi-agent coding agents, agent harness, agent orchestration, persistent agent memory, shared memory for AI agents, asynchronous multi-agent systems, long-horizon agents, inference-time scaling, test-time compute, autonomous research agents, agentic software engineering, Git-based memory, branchable memory, semantic dependencies, multi-agent collaboration, coding agent swarm, program reconstruction, mathematical reasoning agents, negative-result reuse and artifact provenance. October 2026 research roundups show sustained attention to agent harnesses, long-horizon agents, coding-agent workflows and persistent work across sessions. These are topical trend signals from research curation, not measured search volume or a guarantee of ranking.

The practical lesson is to preserve work products, not just conversation history. Make every agent episode produce a small, reviewable artifact: a code change, test, experiment, finding or explicit negative result. Track where it started and what it relied on. Keep branches independent until integration is justified. Make work visible while it is in flight to reduce duplication. Evaluate the final result separately from the accumulation process. Measure how much prior work is reused, how much repository ancestry reaches the selected solution, and how quickly costs grow as memory expands. For long-running agents, durable provenance and reproducible artifacts may matter as much as clever prompting.

GitSwarm reframes inference-time scaling as an information-retention and reuse problem. Its workers do not simply exchange messages; they contribute durable artifacts to a shared, branchable computation graph. The results are promising across program reconstruction, mathematical proofs and small-scale autonomous architecture research. The strongest evidence is not only the headline benchmark gap, but the measured reuse of earlier contributions and the recovery of failed ideas. Important questions remain about causal attribution, memory growth, compute efficiency, correctness and security. The paper's broader insight is compelling: an agent system should be evaluated not only by what one episode produces, but by how much useful capability the entire process accumulates.