Anthropic, Google, and OpenAI researchers collaborated to build a SQLite-like database engine using parallel AI coding agents. The project, written in Rust, generated ~19,000 lines of code across 154 commits, implementing core features like B+ trees, transaction semantics, and query optimization. A key innovation was treating software development as a distributed system, using git for coordination and lock files to manage agent workflows.

The system operated in two phases: a bootstrap stage creating foundational documentation and test harnesses, followed by a worker phase with six agents (two each from Claude, Codex, and Gemini) iterating on implementation. Agents pulled updates, claimed scoped tasks, and validated work against SQLite3 as an oracle. 54.5% of commits addressed coordination challenges like lock conflicts and stale entries, highlighting the critical role of strict task boundaries and anti-entropy measures like unit tests.

While a coalescer agent reduced duplication, it only ran post-project due to mid-experiment failures. The team noted limitations in tracking agent contribution and token usage across platforms. Despite these hurdles, the experiment demonstrated that parallel agents could achieve compounding throughput when given narrow interfaces, common truth sources, and rapid feedback loops.

This approach reimagines software engineering by blending distributed systems principles with AI collaboration. Practical applications include accelerating complex codebases through orchestrated agent teams, though challenges remain in monitoring and resource allocation. The project’s GitHub repository provides a replicable framework for similar experiments.