Meta 论文：GitSwarm 让智能体共享 Git 仓库协作
Meta 把多个智能体丢进同一个 Git 仓库，失败记录也不删，结果在 ProgramBench 上比单跑 Codex 高出 14 个百分点。
Meta 新论文提出 GitSwarm，让多个相同智能体在共享 Git 仓库上并行工作，无需上级分配任务。每个智能体读取历史提交、选择尝试方向，并提交结果及所依赖的早期 commit 列表。在 ProgramBench 的 50 个程序重建任务上，GitSwarm 得分 79.4%，而单个 Codex 智能体在相近算力下最高仅 65.1%。后续智能体有 94.7% 的贡献建立在已保存的提交之上。
New Meta paper shows that AI agents can build on each other's work, even failed attempts, when every step lives in a shared Git repo.
Most ways to give agents more compute treat each run on its own. When a run ends, its partial results and failures vanish, so later runs may need to rediscover them.
GitSwarm runs many identical agents on a shared repo, with no boss assigning tasks. Each agent reads past work, picks what to try, and commits its result with a list of earlier commits it used, even from other branches.
On 50 program-rebuilding tasks from ProgramBench, GitSwarm scored 79.4%, while a single Codex agent told to keep working peaked at 65.1% at similar compute. Later agents built on 94.7% of saved contributions.
Instead of pushing a single agent to keep going, run several over a shared Git history that keeps every attempt, failures included.
