Harnesses Use Guides and Sensors for AI Agents

Birgitta Böckeler (Thoughtworks) describes a practical distinction for engineering AI agent harnesses: guides and sensors. Guides are pre-output constraints—rules files, example code, style guides and repository patterns—that bias an agent toward desired behavior. Sensors are post-output checks—linters, tests, CI gates and audits—that detect failures after the agent produces output. Each approach has predictable failure modes: guides can produce rule fatigue and non-binding suggestions; sensors provide durable enforcement but create slow, expensive feedback loops. Böckeler explains simple diagnostics to identify which side a team underinvests in and recommends a rule of thumb: introduce new constraints as sensors first, then add guides to make correct behavior natural. The piece argues that balanced investment in both guides and sensors is the minimal effective architecture for reliable, scalable agent workflows.