The chat box is a degenerate case of a loop — a loop that runs once and ends. Boris Cherny's point about Claude Code is that the interesting work starts when you stop thinking in chats and start thinking in loops.
A loop is a tick of work. The agent reads, decides, acts, writes, and goes to sleep. The next tick fires on a schedule — a cron job, a queue event, a webhook — and the agent picks up where it left off. The harness around the model is what makes this safe to leave running: explicit permissions, a memory it can read back, an audit log that records every step.
Anthropic uses this pattern internally for what Boris describes as PR-babysitting — scheduled Claude instances watching open pull requests, running the relevant tests, summarising the diff, posting a comment, flagging the ones that need a human. None of it is a chat. All of it is loops.
The other thing scheduling enables is parallelism. A single engineer can have a hundred Claudes running at once, each on its own loop, each on its own branch. The bottleneck stops being the model and becomes the harness — how much work you can responsibly leave running while you sleep.
That is the shape of the studio's AI work. Not chat windows. Loops, on a schedule, with the right room around them.
Filed Mar 2026 · Folkestone