# How Dan Luu actually uses coding agents

danluu.com essays, ex-Centaur

**If you take one thing:** `Named-persona review panels`, with contrarian personas. instead of switching models, he has independent agents review as named engineers (and as himself) plus contrarians, which he says beat the same token budget spent otherwise.

- **Agent:** Codex (cli) ([source, 2026-08-30](https://danluu.com/bug-blind/))
- **Model:** GPT-5.6 Sol ([source, 2026-08-17](https://danluu.com/benchpocalypse/))
- **Review:** Heavy randomized testing instead of human review ([source, 2026-07-03](https://danluu.com/ai-coding/))
- **At once:** 10 agents ([source, 2026-07-03](https://danluu.com/ai-coding/))
- **Harness:** One-off agent-written loop scripts with health checks, on laptops and AWS hosts ([source, 2026-07-03](https://danluu.com/ai-coding/))
- **Rules file:** AGENTS.md (private) ([source, 2026-07-03](https://danluu.com/ai-coding/))

## In their words

> Just like with Google Docs, I consider codex above average in terms of software quality in the space, but even though I haven't been using it for a year, I could easily write 10k words on all the workarounds I've implemented (either by habit or, in some cases, with actual scripts that monitor for broken behavior and then correct it).

> Since my original goal here was to see what happens when you run a current (public) SOTA agent in a loop (GPT-5.6 Sol) without much supervision on a non-trivial code optimization problem without any real supervision

> Just for example, one that I've seen come up repeatedly related to the earlier note that, on tasks that I do, adding a clear, strict, instruction in AGENTS.md to not do something fails once every week or two when juggling 10 agents, a failure rate of something like one per 100 rule-agent-days.

> Personally, I'm very comfortable shipping code without human review because I've seen it done on products that are technically more challenging than most software at most software companies.

> As a result, what I've ended up doing is making one-off loops for tasks that need to run for a while.


Archetype (our label): Orchestrator. Newest source: 2026-08-30.
Card: https://www.prehistoriq.com/stacks/danluu/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/danluu.json