How to use Codex to build a Rust database swarm
Kian Kyars coordinated Codex, Gemini and Claude workers on a Rust SQLite-like engine, using task locks, shared notes and sqlite3 checks while noting limits.
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Build an experimental [systems component] with [agent workers]. Define module interfaces and task locks, require each worker to claim one scoped task, and validate behavior against [reference implementation]. Record progress, merge decisions and unsupported cases; do not present the prototype as production-ready.
Reach people exploring AI agents
See placements →What happened
Engine implements parser/planner/executor/storage and passes a supported subset of SQLLogicTest.
Reported by Kian Kyars on BlogHow the agent did it
- Bootstrap generated skeleton/design/test harness.
- Workers pulled main, claimed scoped tasks with locks, implemented and tested against sqlite3.
- Shared progress/notes and Git pushes coordinated work.
Where a person stepped in
Author set harness/module boundaries and later validated supported SQL against an external suite.
Kian Kyars tested a small heterogeneous coding swarm by asking Claude, Codex and Gemini to build a SQLite-like engine in Rust. Claude bootstrapped the crate skeleton, design notes and test harness. The worker phase then ran two instances of each agent, so this account does not establish that Codex led the build or contributed particular components.
Workers pulled the latest main branch, claimed a scoped task, implemented and tested against sqlite3, updated shared progress and notes, and pushed their changes. Lock files and module boundaries coordinated parser, planner, executor and storage work. Kyars describes task coordination as a substantial part of the commit history.
The prototype implemented a database architecture and passed a limited set of SQLLogicTest queries. It was not a production SQLite replacement: unsupported syntax caused many failures, concurrency was absent and the author identified inefficient buffer cloning and freelist handling. A Gemini coalescer intended to remove duplication stopped midway during the experiment.
The transferable lesson is the harness, with explicit claims, shared state and an external behavior oracle.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · Kian Kyars
Frequently asked questions
How were the workers coordinated?
Each worker pulled main, claimed a task using shared coordination files, implemented and tested, updated notes and progress, then pushed.
Did the engine replace SQLite?
No. It supported only a subset of SQL, lacked concurrency and was explicitly described as an orchestration experiment rather than a production database.
Can the source identify Codex's individual contribution?
No. Two Codex workers participated alongside two Gemini and two Claude workers, but the author could not determine which agent contributed most.
More details
Remaining steps
- Gemini coalescer attempted deduplication.
About this account
Codex workers claimed scoped tasks, implemented and tested against sqlite3, updated shared notes and pushed alongside Gemini and Claude workers.
Primary first-person account opened and checked on 2026-10-14. Not a production database; unsupported SQL and no concurrency. Agent contribution weights not measured. Gemini coalescer stopped midway. Outcomes and test claims are attributed to the source and were not independently reproduced. Prompt is suggested by Agents In Use, not the author's quoted prompt.
Credit: Kian Kyars on Blog
Source published 12 Feb 2026
Source checked 14 Oct 2026
Tools and topics
Codex CLI · Gemini CLI · Claude Code · Rust · Git · sqlite3
Agent coordination · Rust · Database prototype
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.