My agents need three verbs to help me write code: plan, implement, verify. Planning happens locally with the model and my issue tracker. Implementing happens across a fleet of coding agents. Verification has been the painful one. I ended up using GitHub Actions as my trusted environment for testing untrusted code, which works but is miserable ergonomics for an agent: batch execution, a minutes-long feedback loop, and no way to poke at a failure interactively. Unfortunately, verification is the step where the agent needs to iterate quickly.

To me, the interesting part isn't the isolation, RuntimeClass has given us gVisor/Kata pods for years. It makes it easy to deploy this new type of workload where an agent claims a pre-warmed sandbox, gets a prod-like environment with a stable identity, it can hibernate when idle and be thrown away when it has served its purpose.

Replacing CI for the verify step doesn't seem to be the marquee use case, but I'm going to set it up and try it.