Inside the New Agent Swarm: Coordination at 1,000 Commits/Sec
A deep dive into how a new AI agent swarm architecture uses planner-worker hierarchies and a custom VCS to coordinate work at 1,000 commits per second.
An AI research team has refined its approach to coordinating swarms of cooperating AI agents, moving from empirical trial-and-error to deliberate engineering. The new architecture splits tasks into a tree: planner agents run on top-tier models to break down goals, while worker agents on faster, cheaper models execute the leaf-level tasks, improving context efficiency and reducing the drift seen in long-running single agents.
Tested against the old swarm on building SQLite from scratch in Rust using only documentation, the new system passed 80% of a held-out test suite in four hours, while the old swarm had to be paused within two hours. Costs varied significantly across different model pairings even when output quality stayed consistent.
To support hundreds of agents committing simultaneously, the team built a custom version control system from scratch, scaling from roughly 1,000 commits per hour under Git to about 1,000 per second. This exposed new coordination failure modes rarely seen by human teams — split-brain implementations, planner contention, merge conflicts, bloated 'megafiles,' and reluctance to touch core code — each addressed with targeted mechanisms like shared design docs, neutral merge-resolution agents, and licensed breaking changes.
Quality assurance relies on stacked, decorrelated review lenses across different models and context levels, achieving reliability no single reviewer could provide alone. The results offer a clearer picture of how large-scale autonomous software development can be engineered rather than merely discovered.