# How Mitchell Hashimoto actually uses coding agents

Ghostty, HashiCorp co-founder

**If you take one thing:** `Harness engineering`, after every agent mistake. he coined the term. Every bad agent behavior gets an AGENTS.md line or a script so it never happens again.

- **Agent:** Claude app (app) ([source, 2026-08-21](https://x.com/mitchellh/status/2090893729469268109))
- **Agent:** Codex app (app) ([source, 2026-08-21](https://x.com/mitchellh/status/2090893729469268109))
- **Model:** Claude Fable ([source, 2026-09-03](https://github.com/ghostty-org/ghostty/pull/14138))
- **Model:** GPT-6 Astra ([source, 2026-09-09](https://github.com/ghostty-org/ghostty/pull/14189))
- **Review:** Hand-reviews every line; agents validate, he writes all prose ([source, 2026-04-15](https://mitchellh.com/writing/simdutf-no-libcxx))
- **At once:** 2 agents ([source, 2026-08-21](https://x.com/mitchellh/status/2090893729469268109))
- **Harness:** Codex and Claude desktop apps; Ralph loops for background validation ([source, 2026-08-21](https://x.com/mitchellh/status/2090893729469268109))
- **Rules file:** AGENTS.md at root plus 9 per-directory AGENTS.md (Ghostty); each line fixes a past agent mistake ([source, 2026-04-08](https://github.com/ghostty-org/ghostty/blob/9897d6caba05c0cbf256f86bec2e2935f164a9c7/AGENTS.md#L36-L39))
- **Skills:** discovering-upstream-apis, upstream-api-feedback ([source, 2026-04-10](https://github.com/mitchellh/go-libghostty/tree/f79d71fc403d8deec990fdbe4b3a930b87fba98e/.agents/skills))

## In their words

> Then I'd pick up the bug reports, validate them myself, and either fix them myself or kick off new agents manually, just Codex app or Claude app.

> Then I'd pick up the bug reports, validate them myself, and either fix them myself or kick off new agents manually, just Codex app or Claude app.

> **AI usage:** Fable did validation of the work, I did the implementations, commit messages, and PR notes.

> **AI usage:** Fable and Astra both helped with validation, edge cases.

> Fable + Sol (two separate agents) in parallel would write test cases and output them on disk in their own folders.


Archetype (our label): Purist. Newest source: 2026-09-09.
Card: https://www.prehistoriq.com/stacks/mitchellh/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/mitchellh.json