Will Bowers

I work on sandbox infrastructure for AI agents. It’s basically the part that runs untrusted agent code and keeps it from doing anything it shouldn’t. Mostly a containment problem, and most of my time goes into building the system for it.

The agent sandboxes are built on the FaaS platform I’ve worked on for years, which is a ton of short-lived pods under constant churn. Same scheduling and lifecycle problems underneath, just a new workload on top. The trust model is the part that’s different (the FaaS platform isolates tenant code from the host and its neighbors; with agents the workload itself is adversarial, so the sandbox is default-deny and only gets what it needs).