# How Andrew Gallant actually uses coding agents

ripgrep, Rust regex, Jiff

**If you take one thing:** `"I wrote this PR with an LLM"`, flagged for careful review. he says up front which PRs an LLM wrote, reviews them as untrusted and pares them down himself.

- **Agent:** Codex (cli) ([source, 2026-06-03](https://github.com/BurntSushi/ripgrep/pull/3420#issuecomment-4615548108))
- **Model:** GPT-5.6 Sol ([source, 2026-07-30](https://github.com/BurntSushi/dotfiles/blob/1dab3ffec0570fdb2d82ea8411ab07c8ae7f6d95/.codex/config.toml#L3-L8))
- **Review:** Flags his LLM PRs for careful review, then simplifies them by hand; Codex review catches bugs too ([source, 2026-06-07](https://github.com/rust-lang/regex/pull/1364#issuecomment-4642193394))
- **Harness:** Codex in its workspace-write sandbox, approvals on request ([source, 2026-07-30](https://github.com/BurntSushi/dotfiles/blob/1dab3ffec0570fdb2d82ea8411ab07c8ae7f6d95/.codex/config.toml#L1-L2))
- **Hooks:** Codex notify: his own `A` alert script (desktop notification and sound, or a push to his phone when away) ([source, 2026-07-30](https://github.com/BurntSushi/dotfiles/blob/1dab3ffec0570fdb2d82ea8411ab07c8ae7f6d95/.codex/config.toml#L4-L7))

## In their words

> I got this benchmark from Codex that I think I believe.

> model_reasoning_effort = "max"
notify = [
  "sh", "-c",
  "A bart 'Codex' 'Done or waiting for input'",
]
model = "gpt-5.6-sol"

> I wrote this PR with an LLM, so this still needs careful review. From a glance, the solution here seems very over complicated. But it does pass tests and rebar benchmarks.

> approval_policy = "on-request"
sandbox_mode = "workspace-write"

> notify = [
  "sh", "-c",
  "A bart 'Codex' 'Done or waiting for input'",
]


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