On August 19 Simon Willison published smolmachines / smolvm as a sandbox for untrusted Python & JavaScript, a research note on whether smolmachines can serve as a fast, secure sandbox for code you did not write.
The brief he set was specific: put smolmachines.com through its paces as a fast secure sandbox, and work out what it would take to run untrusted Python and JavaScript with limits on RAM and CPU time — protection against the classic while true — and no network access.
The method is as much the story as the result. Willison did not do the exploration himself; he handed the task to Claude Fable 5 running in Claude Code for web and published the output as a research artifact in his simonw/research repo, which is where the write-up and code live.
Cheap, hard-limited sandboxes are the unglamorous bottleneck under every code-executing agent: if the model can run arbitrary Python, something has to bound the memory, the CPU and the network. Willison's delegate-the-research-to-an-agent-and-publish-the-repo pattern is itself becoming a recognisable format for evaluating this class of tooling.
Sources: Simon Willison, simonw/research on GitHub