# How Terence Tao actually uses coding agents

Fields Medal mathematician, UCLA

**If you take one thing:** `Five favorable conditions`, before handing code to an agent. he delegates all the coding only when the work is non-critical, stand-alone, sandboxed, off his core skills and not competing with humans.

- **Agent:** Claude Code ([source, 2026-05-04](https://mathstodon.xyz/@tao/116517314420904580))
- **Model:** Claude Opus 5 ([source, 2026-10-03](https://github.com/teorth/tao-web/commit/90fdcfc719e65aba141c75937e7168fe610f4f7a))
- **Review:** Checks results, doesn't read machine code ([source, 2026-09-25](https://github.com/teorth/IEANTN/blob/0f4454a3bdc92e85cd82722724b68c80fd14cb55/CLAUDE.md#L89-L91))
- **Harness:** Spec-first recipe per app, background sub-agents ([source, 2026-07-14](https://teorth.github.io/tao-web/apps/porting-legacy-applets.html))
- **Rules file:** CLAUDE.md for the IEANTN project (94 lines); AGENTS.md points to it ([source, 2026-09-25](https://github.com/teorth/IEANTN/blob/0f4454a3bdc92e85cd82722724b68c80fd14cb55/CLAUDE.md#L41-L42))

## In their words

> This time around, I uploaded the report to the Claude Code agent, which was able to inspect the report, the LaTeX source, and the PDF version of the paper, identify the corrections, and propose unambiguous fixes for eleven of the twelve corrections

> Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

> Solutions are exempt: they are machine-generated, verified once, and never read.

> Unlike the original apps, most ports were driven by a spec-first, per-app recipe run partly through background sub-agents; the shared method matters more than any single exchange.

> **Never weaken a statement to make a proof work.** The source paper is ground truth. If the
  stated bound is wrong, say so — do not quietly adjust it.


Archetype (our label): Skeptic. Newest source: 2026-10-03.
Card: https://www.prehistoriq.com/stacks/teorth/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/teorth.json