Four checkpoints before an agent-built feature ships

Checkpoint 01 — scope

Write the one user action the feature must make easier. Cut everything that does not support it.

Checkpoint 02 — vertical slice

Build the smallest end-to-end path: real input, persistence, empty state, and error state.

Checkpoint 03 — evidence

Run tests, open the route in a browser, and capture the happy path plus one failure state.

Checkpoint 04 — handoff

Ask the agent to summarize changed files, known risks, rollback steps, and the command that proves the feature works.

Rule

Speed is useful only when the loop produces something another person can inspect and trust.

Research note

The emphasis on agents running tests and capturing screenshots is adapted from @kaxil: https://x.com/kaxil/status/2037503513350005134