A CHI 2026 study asked senior engineers how they work with an AI agent. They did not hand it the problem. They scoped the change, explained it in detail, and delegated small units like boilerplate and tests, then insisted on changes small enough to review.
Control here is not a feeling. It is a size. A piece you can read in one sitting is a piece you can judge.
You do not stay in charge of an agent by trusting it less. You stay in charge by keeping each handoff small enough to check.
Kay
P.S. Course 0 is where these ideas become practice: https://systemthinkinglab.ai/course-0