Agents
Closed Loop
An open workflow ends with the result: the system delivers, someone signs off, done. A closed loop leads back: what happened in operation is measured, compared with what should have happened, and the deviation changes the system. The difference sounds academic and decides durability. Models that have learned from examples lose their fit as soon as reality drifts away from the examples: new products, new customers, changed regulations, different language use. This is not a defect but the normal case, and it happens slowly enough that nobody notices. That is why classic acceptance doesn't hold up here: check once and approve was meant for products that don't change.
An example
A classification system runs unchanged for a year and gets worse without anyone noticing: the product range has changed, so has the customer base, the categories no longer fit. A closed loop would have caught it: the clerks' corrections flow back into the test set, the hit rate is measured weekly, and above a threshold someone is notified. The difference is not the technology but that the measurement takes place at all.
A closer look
The loop has four stations: act, measure, evaluate, adjust. The second is decisive. Without a measurement that was fixed before the action, the loop is open and the evaluation is after-the-fact interpretation.
Why it matters
Systems that have learned from examples lose their fit when reality drifts away from the examples. This is not a malfunction but the normal case, and the reason classic acceptance on the pattern of “check once, then approve” doesn't hold up here. A closed loop needs four things: ongoing measurement, a threshold with an addressee, a path from the correction back into the test set, and someone authorized to stop the system. The fourth is the one most often missing.
Commonly misunderstood
Most “closed loops” never close. The feedback is missing: results are reported but not fed into the next round.
Where this term does its work in the magazine.
2 articles use the term or take up its principle.
The Long View
Who's Actually Leading Whom?
The more autonomous a system appears, the earlier responsibility has to be designed.Read the articleFrom Science