Correcting course mid-task
A session doesn't have to run to completion before you weigh in. If you watch the agent head toward an approach you know is wrong — say, it's about to add a new dependency when the codebase already has a utility that does the same thing — the effective move is to interrupt and say so immediately, not wait for it to finish and then request a rewrite. The agent incorporates your correction into its next action the same way it incorporates the result of a failed test: as new information to act on, not as a nuisance to route around.
Course correction works best when it's specific. "That's wrong, try again" gives the agent as little to work with as a vague initial task did. "Use the existing formatCurrency helper in utils/money.ts instead of adding a new library" gives it an exact, actionable substitution. The clearer the correction, the fewer additional round trips it takes to get back on track.