Software & DevOpsDone with custom-built or unnamed agent Feb 16, 2026
Build SQLite with a swarm of parallel coding agents
The author recounts using parallel coding agents to build a SQLite implementation and shares experience from it.
- What They Did
- One Claude run bootstrapped docs, a crate skeleton and a test harness. Then six workers (two each of Claude, Codex and Gemini) looped in separate workspaces. Each pulled main, claimed a task through a lock file, implemented it and tested against sqlite3 as an oracle, updated shared notes, and pushed. A separate Gemini coalescer was meant to remove duplication.
- How It Went
- About 19k lines of Rust in 154 commits, with 282 AI-written unit tests passing. It passed only 64 SQLLogicTest queries, since subqueries and CASE are unsupported. It has no concurrency, and the code has some slop.
- Worth Knowing
- 54.5% of commits were lock and claim coordination. The coalescer ran only once, at the end, and Codex used 100% of a Pro weekly quota.
Try it yourself with custom-built or unnamed agent
I want to build a small [project, e.g. SQLite-style database engine] in [language] using several parallel agents. First write a spec that splits the work into independent modules ([e.g. parser, storage, query executor]), then run one sub-agent per module in separate worktrees with tests for each. Stop when the modules are merged and a test suite of [number] basic queries passes, and show me what failed along the way.