Posts here

An agent can fix the code, write the tests, and still preserve the wrong picture of the real app. In one maintained personal app, green tests survived three attempted fixes until a human tried the behavior on a desktop; the repair needed a check against the real browser, not another green suite.