# How David Crawshaw actually uses coding agents

exe.dev, Tailscale co-founder

**If you take one thing:** `meat`, strips diffs to what matters. his own LLM tool that cuts imports, nil-checks and error handling out of a diff, so his review covers only the architecture.

- **Agent:** Shelley (web) ([source, 2026-08-02](https://blog.exe.dev/devtools-must-be-open-source))
- **Model:** OpenAI models ([source, 2026-08-03](https://news.ycombinator.com/item?id=49157584))
- **Review:** Reads agent code before it hits serious systems, trimmed by meat ([source, 2026-08-02](https://blog.exe.dev/devtools-must-be-open-source))
- **Harness:** Fresh, unconstrained VM per agent (exe.dev) instead of built-in sandboxes ([source, 2026-02-08](https://crawshaw.io/blog/eight-more-months-of-agents))

## In their words

> I like this tool, but it has two downsides: first, I like to read my diffs in Shelley with a good UI, not in a terminal.

> I still use oai as my daily but I have been very impressed by qwen/glm/k3/etc when I try them.

> The principle is that while agents write code, I still read it before pushing to our serious systems.

> You have to turn off the sandbox, which means you have to provide your own sandbox. I have tried just about everything and I highly recommend: use a fresh VM.

> Humans need to review agent-written code in critical systems.
But models are good now. You don't need to review for style or nil-checks
or imports. You need to review concepts, algorithm choices, architecture.

So meat uses a model to reduce a diff to the important parts.
It shows you the meat.


Archetype (our label): Toolsmith. Newest source: 2026-08-03.
Card: https://www.prehistoriq.com/stacks/crawshaw/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/crawshaw.json