I wonder how long it will take for someone to pwn this?

We've found that methods like this substantially increase the quality and reliability of coding agent output. The ability to run code in a sandbox, drive an interactive session using a browser or API calls or other apps, and visually confirm output via vision models all adds up to plugging a big hole in the feedback loop for agent modifying a complex codebase.

We've had agents go as far as interactively testing how our product responds in video calls by launching our full stack in a set of docker containers (app, api, db, queues, etc.), all inside a larger sandbox, populating test data, connecting the mock system to a real video call solution like Google meet, and injecting audio and video to test the response. End-to-end, like a real user flow.

It's not perfect yet, but if you are a skeptic on the ability for AI agents to productively modify a complex product, I'd highly encourage you to play with a setup like this before ossifying your conclusions.

How does this work when a projet have many external dependencies, like an S3 bucket, a secret manager, a third party API, etc?