When you repeatedly tell an AI what to do next, it is natural to ask whether the next instruction could be chosen from the previous result. Feedback offers a useful starting point. This five-part series connects physical control loops with AI workflows while keeping their differences clear.
Two meanings of loop
Control engineering studies systems that feed measurements back into decisions about physical actions. In AI work, loop engineering describes designing the system that discovers work, executes it, checks results, and chooses what happens next. Addy Osmani's June 2026 essay presents the latter perspective. This series uses it as a starting point, rather than treating the emerging phrase as an established replacement for control engineering. Source: Addy Osmani
Four roles in temperature control
Consider an illustrative room-temperature target of 24°C and a measurement of 22°C. A controller decides how much heat to request. The heater changes the room, and a later measurement informs another decision. The roles are target, observation, decision, and action.

Fixed heater output cannot use the measured effect of an open window to correct its action. Closing the feedback path helps, but a faulty sensor can still lead to inappropriate actions. Feedback does not guarantee good control.
Where these loops are used
NI lists temperature, flow, speed, pressure, and voltage as PID control variables. Thinking about the measured quantity and the available actuator makes these applications concrete: temperature and heater output, for example. This is a vendor's explanation of application categories, not a count of deployments. Source: NI
For AI, Anthropic's 2025 Research system is a documented product example: a lead agent plans a research process and delegates searches. New findings can change the direction of investigation. The later phrase loop engineering should not be retroactively presented as the original product name. Source: Anthropic
An AI score is not a thermometer
For writing, an objective might be an explanation a reader can follow. Actions are edits; observations might include missing-step checks or a reader attempting the instructions. A score of 90 alone does not explain what improved.

Physical measurements also contain error, but AI tasks may have ambiguous evaluation criteria as well. Edits can invalidate previously passing checks. A PID equation cannot simply be applied to prose quality.
Exercise: identify the returning signal
A: summarize the same document every morning and save it. B: compare numbers in the summary with the original, repair mismatches, and check again. Which changes its behavior in response to results?
B does. A is scheduled repetition, but no result-dependent change is specified. B is still limited: checking numbers does not establish that the meaning is correct. Pick one repeated task of your own and write down its target, observation, decision, and action. If observation only says that the AI announced completion, identify something in the artifact you can check directly.
Read the five-part series
- Loop Engineering 1: Feedback, from Temperature Control to AI (this lesson)
- Loop Engineering 2: Error, PID, Delay, and Oscillation
- Loop Engineering 3: Design an AI Action–Evaluation Loop
- Loop Engineering 4: Failure, Budgets, and Safe Resumption
- Loop Engineering 5: Design and Test Your First Small Loop











