Agents need executable feedback. Have them write the test first, watch it fail, implement the change, and watch it pass. Use tight executable feedback loops to steer agent work.