# How José Valim actually uses coding agents

Creator of Elixir, Tidewave

**If you take one thing:** `Adversarial review agents`, on compiler & security work. he makes a second set of agents try to disprove a change before any human sees it.

- **Agent:** Codex ([source, 2026-07-31](https://x.com/josevalim/status/2083202612162077123))
- **Agent:** Tidewave (app) ([source, 2026-08-28](https://x.com/josevalim/status/2093304967360548895))
- **Model:** Sol 5.6 ([source, 2026-07-17](https://x.com/josevalim/status/2078177113560244355))
- **Review:** Reviews agent PRs like any other code ([source, 2026-09-06](https://x.com/josevalim/status/2096683774662611162))
- **Harness:** Tidewave's live code-review pane, marking hunks as reviewed ([source, 2026-04-15](https://tidewave.ai/blog/code-review))
- **Rules file:** AGENTS.md symlinked to CONTRIBUTING.md in Elixir; own AGENTS.md in Lynx ([source, 2026-06-15](https://github.com/elixir-lang/elixir/commit/1d599978f7d36ea6896b294b3d26d110212be7ab))

## In their words

> It is the reason why I prefer Codex, which bundles fast mode in the subscription, compared to Claude Code, which bundles larger context windows.

> Tidewave now ships with a design canvas. It has been so useful internally that we're kicking ourselves for not shipping it earlier!

> You probably heard me saying "coding agents are useless for type system work 🥴" but Sol is definitely an improvement.

> Outside of that, I mainly ask the agent to break any large PR into small ones. Then I review them like any other code 😅

> Tidewave updates the code review pane in real-time and allows you to mark which code sections (hunks in git terms) have been reviewed


Archetype (our label): Rules Lawyer. Newest source: 2026-09-06.
Card: https://www.prehistoriq.com/stacks/josevalim/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/josevalim.json