AI · November 18, 2026 · Low impact
Simon Willison and the Prime Radiant lab introduce an external sandbox pattern for coding agents
The pattern runs code executed by an agent inside a dedicated sandbox while the agent itself lives outside that sandbox. The goal is to make it safer to try local models without a mistake or prompt injection causing damaging actions on a machine. The work was done with Prime Radiant, the AI research lab behind Superpowers and other projects. The write-up includes a demonstration of the approach.
Separating the agent from the environment where its code runs is a practical defense for anyone letting agents execute code on their own machines.