# How Thorsten Ball actually uses coding agents

Amp co-founder, interpreter book

**If you take one thing:** `“Irrefutable proof”`, video proof, not line reads. rather than read the diff, he asks the agent in its orb to prove the work end to end, often with a video.

- **Agent:** Amp (web) ([source, 2026-09-20](https://registerspill.thorstenball.com/p/joy-and-curiosity-100))
- **Model:** Claude Opus 5.5 ([source, 2026-09-29](https://x.com/thorstenball/status/2104860291674886642))
- **Model:** GPT-6 Astra ([source, 2026-09-12](https://registerspill.thorstenball.com/p/joy-and-curiosity-99))
- **Review:** Spot checks and architecture, not every line ([source, 2026-08-16](https://registerspill.thorstenball.com/p/joy-and-curiosity-95))
- **Harness:** Amp agents in remote orbs, no local dev environment ([source, 2026-08-16](https://registerspill.thorstenball.com/p/joy-and-curiosity-95))
- **Rules file:** AGENT.md in his dotfiles (26 lines) ([source, 2025-05-26](https://github.com/mrnugget/dotfiles/blob/72270ffcb997eafeeb46112f5a12adcdf2493afa/AGENT.md#L1-L4))

## In their words

> I had Amp build a little Copilot-style autocomplete for a shell, with Jev picking the next most likely command from shell history.

> Opus 5.5 did a great job. It feels nice to use and all the keybindings you expect work.

> Something changed with these latest models, with Fable 5.1 and GPT-6 Astra.

> And for all of the things I shipped here, I didn’t review each line anyway. I do spot checks and make sure the architecture is right, yes, but do I need local tools for that? No.

> I have not used my local development environment for any of this. I’ve done all of this remotely, using Amp, in orbs.


Archetype (our label): Orchestrator. Newest source: 2026-09-29.
Card: https://www.prehistoriq.com/stacks/mrnugget/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/mrnugget.json