Your agents need a sandbox, not a desert

Everyone agrees agents need code execution. The agreement ends the moment you ask how to do it. The usual answer is "my agent needs a full Linux VM to succeed." That answer suits sandbox providers, and it's often wrong.

Plenty of real-world agent workflows don't need a whole computer. No arbitrary packages, no shell access, no CPython or node, never mind awk, sed, and gcc. The model needs a small amount of safe, expressive compute: enough to write code, call tools, and keep intermediate state out of the context window. That's the idea behind Pydantic Monty, a minimal Python interpreter written in Rust and built for running code that agents write.

Samuel Colvin, founder and CEO of Pydantic, makes the case that for a large class of agent systems, a curated set of tools in a custom runtime beats a full sandbox. Not because full sandboxes are bad, but because they solve a much bigger problem than most embedded agents have, and you pay for the mismatch in complexity, cost, operational pain, and latency that can run 100,000 times higher. There's such a thing as too much sand. Often the limits of a small, custom-built sandbox are a feature, not a bug.

A 30-minute session, with time for your questions at the end.