alt Hacker NewsQuite the opposite, AI should work longer and interrupt the human less often (Tokens are cheap, interruptions are expensive). So we want to push the agent's horizon to infinite, now when they do make interruptions (e.g. Creating diff) it will be larger chunks and more complex, so these summaries are actually quite useful.
You can and should have both smaller chunks and a larger time horizon. The AI should output code in a format that's easy to review.
The Linux kernel submission guidelines are one thing you can feed the AI to guide that. Work is submitted as a patch set, and each patch in the set must be small and self contained. Many patch sets are over 50 patches.