Open vs Closed Agent Loops
Closed loops have clear validation and stop rules. Open loops explore uncertain tasks and need stronger human supervision.
The examples below apply one rule: autonomy rises only when success can be checked independently and safely.
Why it matters
Knowing which kind of loop you have tells you how much autonomy is safe. A closed loop with a crisp validator can run semi-autonomously; an open, exploratory loop should keep a human closely in the loop.
Turn uncertain work into a bounded review
- 01Name the decision
- 02Separate checks from judgment
- 03Add limits and exits
- 04Run one small case
- 05Review the evidence
Do not force an entire exploratory task into a fake pass/fail loop. Close only the parts that can be checked: source retrieval, link validation, build output, or a fixed comparison table. Pause before the remaining judgment, show the evidence and limitations, and let a person choose the trade-off.
Open, closed, or supervised?
| Signal | Closed loop | Supervised loop | Open work |
|---|---|---|---|
| Success test | Objective and repeatable | Partly objective, partly judgment | No decisive test yet |
| Safe next action | Run inside fixed limits | Automate checks, pause for judgment | Research and present options |
| Human role | Approve the final change | Review at named decision gates | Own the decision and trade-offs |
| Stop condition | Check passes or budget expires | Evidence conflicts or a gate is reached | Time box ends or evidence is sufficient |
Choose the smallest safe autonomy level
- 01Can an independent check prove the requested outcome?Yes: start with a closed loop
- 02Can permissions, files, cost, and retries be bounded before the run?No: keep human control
- 03Does any step require product, legal, security, or editorial judgment?Yes: add a named approval gate
- 04Would a green check still be misleading after a shortcut?Yes: add evidence and a separate checker
Practical checklist
- Ask whether success is objectively checkable
- Treat exploratory tasks as open loops
- Increase supervision as uncertainty rises
- Convert open loops to closed ones by adding a validator where possible
A safe example with explicit exits
A documentation refresh can close link checks, spelling, build output, and metadata validation while leaving accuracy and usefulness as a supervised editorial decision.
Stop rule
Stop after two failed validation attempts, any unsupported factual claim, or a requested edit outside the approved documentation paths.
Fallback
Keep the verified source notes and produce a short list of unresolved claims, affected pages, and the exact evidence still needed.
Human approval
A person reviews whether the revised explanation is accurate, genuinely useful, and appropriate to publish; passing automated checks never grants publication authority on its own.
Common failure modes
Practical evidence
Three loop classifications
The exact command, expected exit status, scoped diff, and approval gate make success objectively checkable.
Evidence can inform the decision, but no machine check can decide which trade-off the organization should accept.
Link, build, and metadata checks can close part of the loop; usefulness, claims, and publication still require editorial judgment.