Agent Atlas

The Builders

Peter Steinberger

Ships agent CLIs and Mac apps, and writes the workflow down on his own site.

Stay on main, and leave the model on high.

Steal this

  • Stay on main, and leave the model on high.
  • Keep docs in the repo, and point one global file at them.
  • Budget real time for refactoring. His was about a fifth.

The lesson and the takeaways are our summary. Everything below is theirs, with where they said it.

The setup

Factual
Machines

A MacBook Pro on one screen and a Jump Desktop session to a Mac Studio on another.

Software

Codex CLI, a global AGENTS file, per-project docs, and Tailscale between the Macs.

Models

gpt-5.2-codex at high reasoning effort. The 14 Oct 2025 post, which this one says he has not changed much from, used gpt-5-codex on mid.

As of December 28, 2025. Setups change; follow the links.

Their gear

Mac Studio

Mac Studio

A bigger desktop Mac. Agents run on one for Emanuel; Steinberger reaches his over Jump Desktop.

From
$2,499 (Apple; Apple)
Base
M5 Max, 18-core CPU, 36 GB
Runs
10 or more agents, and mid-size local models.
Maker's page →Photo: Yasu, CC BY-SA 3.0
MacBook Pro

MacBook Pro

Simon Willison's 128GB M5 Max runs local models. Steinberger works on one beside the Studio.

From
$1,999 (14-inch) (Apple)
Base
M6, 16 GB, up to 32 GB
Runs
4–6 agents while it is open and awake.
Maker's page →Photo: AzureSaturn, CC0

Links go to the maker's own page. We earn nothing from them. Every machine is on the gear shelf.

How they work

Factual
  1. On 28 Dec 2025 the workflow is still close to the October one. (Peter Steinberger)
  2. He works on a MacBook Pro and a Jump Desktop session to a Mac Studio at the same time. (Peter Steinberger)
  3. The go-to model is gpt-5.2-codex high. He does not bother with xhigh. (Peter Steinberger)
  4. He is usually the bottleneck. He commits to main and no longer restarts the session for each new task. (Peter Steinberger)
  5. Docs live in the repo. A global AGENTS file points at them and at the Tailscale network. (Peter Steinberger)
  6. On 14 Oct 2025 he ran 3 to 8 agents in one terminal grid, mostly in the same folder. (Peter Steinberger)
  7. The agent file was about 800 lines, written by Codex, and he spent about 20 percent of his time on refactoring. (Peter Steinberger)
  8. Four OpenAI subscriptions and one Anthropic subscription were about 1k a month. (Peter Steinberger)

Where they talk and write

Interviews 34

28 older

Writing 36

30 older

Each link goes to the original page, newest first.