Loop Engineering recommends: At the end of each day, let the agent do an architecture physical examination.
The original prompt word of the architecture satisfaction loop is very short: refactor until you are happy with the architecture.
This sentence is a bit dangerous to read alone, but the version in the Loop Library adds several key guardrails: small step refactor, live test, autoreview, checkpoint commit, and recording progress.
It’s better suited as a “closing move” at the end of a day of development.
During the day, many local decisions were made for the sake of speed.
At night, the agent was asked to look back at duplication, naming, module boundaries, and coupling, and clear the obvious architectural debt first.
The most important thing here is rhythm.
Don’t let the agent make major changes in one go.
Only move a small piece at a time, run tests, do reviews, roll back, and continue with the next round.
Harness and verifier are key here: the agent can be allowed to do more, but every step must have evidence, boundaries, and exit paths.
The video starts here: Corresponding loop: Open source project:
