# How Chip Huyen actually uses coding agents

AI Engineering author

**If you take one thing:** `Cross-provider runner`, Codex, Claude spawn each other. she built her own runner so a strong model plans and reviews while Codex and Claude agents hand work to each other.

- **Agent:** Codex ([source, 2026-07-30](https://github.com/hugobowne/show-us-your-agent-skills/blob/2c1f52025a1a993a5ce58c744d6a6e987ff58c2e/episode-field-notes/ep-8/chip.md#L113))
- **Agent:** Claude Code ([source, 2026-07-30](https://github.com/hugobowne/show-us-your-agent-skills/blob/2c1f52025a1a993a5ce58c744d6a6e987ff58c2e/episode-field-notes/ep-8/chip.md#L117-L119))
- **Model:** Claude Fable ([source, 2026-07-30](https://www.youtube.com/watch?v=NH-ic7-V-jY&t=3048))
- **Review:** A second agent reviews the plan; UI work comes with screenshot evidence on the PR ([source, 2026-07-30](https://github.com/hugobowne/show-us-your-agent-skills/blob/2c1f52025a1a993a5ce58c744d6a6e987ff58c2e/episode-field-notes/ep-8/chip.md#L50))
- **At once:** 10 subagents per lead ([source, 2026-07-30](https://github.com/hugobowne/show-us-your-agent-skills/blob/2c1f52025a1a993a5ce58c744d6a6e987ff58c2e/episode-field-notes/ep-8/chip.md#L96))
- **Harness:** Her own cross-provider runner: a strong agent plans and reviews, cheap agents implement, each in a git worktree ([source, 2026-07-30](https://github.com/hugobowne/show-us-your-agent-skills/blob/2c1f52025a1a993a5ce58c744d6a6e987ff58c2e/episode-field-notes/ep-8/chip.md#L34))
- **Skills:** taxonomy labeling, verified front-end, write tests, author skills ([source, 2026-07-30](https://github.com/hugobowne/show-us-your-agent-skills/blob/2c1f52025a1a993a5ce58c744d6a6e987ff58c2e/episode-field-notes/ep-8/chip.md#L58-L80))

## In their words

> Desktop usually hides the trace of all the commands that it runs, whereas the CLI is more verbose, but I can see a lot more things at a quick glance.

> I can just go out, and I can close my laptop, and it works, which is amazing.

> I feel like it's very wasteful to use uh strong models like fable to do actual implementations.

> With my runner, there's no limit, so it can spin up 10 sub-agents.

> I give it a process: do this, and then get another to review it, and then do post-mortems.


Archetype (our label): Swarm Lord. Newest source: 2026-07-30.
Card: https://www.prehistoriq.com/stacks/chiphuyen/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/chiphuyen.json