Cornelia Davis

AI agents are already well entrenched in the traditional software delivery lifecycle: they build, review, test, and deploy code, yet even as we automate more and more, we are still talking about a controlled build-test-deploy cycle.

But what happens when an agent writes and executes new code not as a deliverable, but as a tool for accomplishing its current goal? There is no pull request, deployment pipeline, or human checkpoint. This is testing in production in its most literal form.

The DevOps movement made testing in production viable by creating safety nets: containers, progressive delivery, observability, and rapid rollback. Runtime code generation by autonomous agents needs its own safety net, and a sandbox is only the beginning. A sandbox controls where code runs; it does not fully control what that code can do.

In this demo-heavy session, we’ll examine the execution model required for agents to safely create and run code without human intervention. Generated code must remain inside the host agent’s governed context: inheriting its tool permissions, policy constraints, and relevant event history, without acquiring capabilities the agent itself does not possess.

We’ll introduce an execution environment built around these principles, explore its architecture, and share practical patterns for making runtime code generation observable, auditable, and bounded.