Build SQLite with a swarm of parallel coding agents
tl;dr. Task: The author orchestrated six parallel coding agents (two each of Claude, Codex, and Gemini) to build a ~19k-line Rust SQLite-like engine. The workflow involved a bootstrap phase generating scaffolding, followed by a worker loop where agents claimed scoped tasks, implemented features, and tested against sqlite3 as an oracle. Human involvement was limited to initial setup and monitoring, as the source does not detail ongoing manual intervention during the agent loop.; tools: custom-built or unnamed agent. Human role: The original post does not say how a person stayed involved.; limitations: The project is a 'simulacrum' of SQLite, not a production database, lacking concurrency and full SQL support. High coordination overhead and documentation bloat (PROGRESS.md reached 490 lines) hindered efficiency. Token usage tracking is difficult with CLI agents, and the coalescing agent failed to complete deduplication tasks during the run..
Tools
custom-built or unnamed agent
Published on Agent Work Atlas:
Overview
tl;dr
Workflow
The author orchestrated six parallel coding agents (two each of Claude, Codex, and Gemini) to build a ~19k-line Rust SQLite-like engine. The workflow involved a bootstrap phase generating scaffolding, followed by a worker loop where agents claimed scoped tasks, implemented features, and tested against sqlite3 as an oracle. Human involvement was limited to initial setup and monitoring, as the source does not detail ongoing manual intervention during the agent loop.
Reported result
The engine passed 64 SQLLogicTest queries but failed most others due to unsupported syntax like subqueries. A significant coordination tax was observed: 54.5% of commits were for lock management rather than feature development. The author exhausted 100% of Codex Pro weekly usage and 70% of Claude Pro weekly usage during the experiment.
Human role
The original post does not say how a person stayed involved.
Limitations
The project is a 'simulacrum' of SQLite, not a production database, lacking concurrency and full SQL support. High coordination overhead and documentation bloat (PROGRESS.md reached 490 lines) hindered efficiency. Token usage tracking is difficult with CLI agents, and the coalescing agent failed to complete deduplication tasks during the run.
Try it
I want to replicate a parallel agent coding experiment for a small Rust project. First, set up a git repository with a lock-file-based task claiming system. Install CLIs for Claude, Codex, and Gemini. Create a bootstrap script that generates a Cargo.toml, DESIGN.md, and PROGRESS.md. Then, launch two instances of each agent in separate terminal panes using a loop script that pulls latest main, claims a task from a 'current_tasks' directory, implements code, runs 'cargo test', and pushes. Monitor the git log for coordination commits vs feature commits. Before running the full swarm, execute a dry run with one agent on a trivial task to verify the lock mechanism works. Ask me for confirmation before spending API credits or pushing to a shared remote. Report the ratio of coordination commits to total commits after 10 iterations.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.