Agent Atlas

The Builders

Simon Willison

Co-creator of Django, maker of Datasette and the llm command.

Context is something you can print and read before the model ever sees it.

Steal this

  • Print the context and read it before the model does.
  • Give a second agent its own fresh checkout.
  • Try local models on your own hardware, and write down the speed.

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

The setup

Factual
Machines

A 128GB M5 Max MacBook Pro, and an NVIDIA DGX Spark.

Software

LM Studio on both. Pi on the Spark, pointed at that LM Studio model.

Models

Qwen 3.8 27B, the 17GB Q4_K_M build. Around 15-30 tokens a second in LM Studio, his figure.

As of August 16, 2026. Setups change; follow the links.

Their gear

NVIDIA DGX Spark

NVIDIA DGX Spark

A desktop box built for running models. Simon Willison drives Pi with a local model on his.

From
$4,699 list; $6,949.99 at a retailer in Oct (NVIDIA Developer Forums; Micro Center)
Base
20-core Arm, 128 GB
Runs
Built to run models locally. Overkill as an agent box.
Maker's page →Photo: Daniel Lu, CC BY-SA 4.0
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
  • Claude Code Anthropic's coding agent.
  • Codex OpenAI's coding agent.
  • files-to-prompt Turns a folder into one prompt you can read.
  • LM Studio Download and run open models on your own machine.
  • Pi A small agent harness you can read and change.

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. files-to-prompt turns a folder into one prompt you can inspect. (Simon Willison)
  2. On 5 Oct 2025 his daily drivers were Claude Code on Sonnet 4.5, Codex CLI on GPT-5-Codex, and Codex Cloud from his phone. (Simon Willison)
  3. He had not adopted git worktrees. Two agents on one repo meant a fresh checkout, often in /tmp. (Simon Willison)
  4. On 16 Aug 2026 he got around 15-30 tokens a second from Qwen 3.8 27B in LM Studio. (Simon Willison)
  5. The same local model drove the Pi harness on the Spark. (Simon Willison)
  6. On 29 Oct 2026 he said he pays for subscriptions with both OpenAI and Anthropic. (Simon Willison)

Where they talk and write

Interviews 53

47 older

Writing 40

34 older

Each link goes to the original page, newest first.

Habits worth copying