Comment by deet
5 hours ago
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.
Didn't Claude Fable do this? (and I think codex and Claude Code in general)
When Fable was around last week, I was smitten with it. I took an executable file from an old DOS application, told it to port it to the Mac. From that single prompt, it was able to set up a test rig with Dosbox to execute the application after already disassembling and gathering as much info as it can and then continuously refine the output application while testing it against the original file. 15 minutes later it had an 99% identical looking and functioning application running natively on the Mac. Sone final refinements got that to 100%.