R1 · how it works · coding agent

A backend feature, end to end.

One small feature is enough to show the whole loop: plan output, code changes, review evidence, and why the final human review is shorter than it would be otherwise.

Take a routine task such as adding rate-limit middleware to an authentication endpoint. The plan explains which files change, where the risk sits, what tests need to exist, and which rollback concerns apply before execution starts.

Execution then happens in an isolated worktree. The receipt accumulates code reads, writes, and verification runs. If review catches a regression, the retry is not mysterious; it is a second attempt attached to the same causal chain.

What the human reviewer gets at the end is not “trust me, it is done.” It is the diff, the gate results, the review outcome, and the narrow places still worth reading slowly.

See the live coding comparison on the homepage.