Martin Keen maps four agent loops in this nine-minute IBM Technology explainer, using library upgrades as a running example.
The shared cycle
- A goal leads to an action, an observation of the result, and an adjustment before the next action.
- Autonomy depends on who supplies the goal, starts the run, decides it has finished and reviews the output.
1. Interactive: a person judges each turn
- A human gives the prompt, inspects the result and decides whether to ask for another iteration.
- “Upgrade library X” can finish when the person accepts the change.
- Fuzzy tasks suit this arrangement because a person can judge an evolving finish line.
- Repetitive work creates a review bottleneck.
2. Delegated: specify the stopping condition
- A human supplies both the goal and a completion rule: upgrade outdated libraries and stop when the build turns green.
- An AI check decides whether to continue the loop or hand back the completed work.
- The human reviews the final result after the agent finishes its iterations.
3. Scheduled: a clock starts recurring work
- Write the goal once and run it nightly, retaining the completion check and human review.
- This suits work that becomes stale over time, such as dependencies that need regular upgrades.
- The human still adjusts or cancels the schedule when its results stop being useful.
4. Event-driven: review by exception
- An event, such as a newly opened pull request, starts a run tailored to that event.
- A separate reviewer agent checks the output with fresh context, potentially using a different model.
- Approval ends the run; requested changes restart it; uncertainty escalates to a human.
- Human review shifts to exceptions in Keen’s most autonomous example.
Caveat: these are teaching patterns. An event trigger alone says nothing about whether human approval is required, and a green build checks only part of correctness.
“Who sets the goal? And who decides when the loop’s done?” — Martin Keen, 1:10