The emphasis on one-shot execution is interesting. Most agent frameworks still rely on iterative loops with human checkpoints, but Stripe's approach of giving the agent a complete context dump upfront and letting it run seems closer to how senior engineers actually work - you read the whole PR/spec first, then write the code. The tricky part is always the context window: once your codebase exceeds what fits in context, one-shot falls apart and you're back to chunked reasoning. Curious if they hit that wall and how they handle repo-scale tasks.

My experiments using a more hands off approach of prompting claude code have always resulted in a two-steps-forward-one-step-back play, where the agent clearly did some good stuff, but did some other stuff in a somewhat undesirable manner, which subsequently needed correcting.

This usually results in A: creating commits where tons of code is being constantly added and removed, B: due to Claude's somewhat cavalier attitude to existing code, has steadily eroded my familiarity with the code base.

I'm still not convinced that these longer loops are that beneficial, compared to 1min prompts to 5-10min AI work.

Doesn't delegating make this a lot more possible? You can fire off a request to a sub-agent, they respond with some predictable status that you can parse, and then you continue (you being the "main" agent), so the context window can remain relatively small. Kinda like how a human does it.