Should an AI agent keep going until the job is done?
An AI agent should usually keep going only within limits a person sets before the run begins. A single pass is enough for a lookup; an objective may need rounds. The stop belongs to the person, not the model, because the model cannot tell whether a marginal step is worth your time, money or attention.
What is the difference between a one-pass run and working in rounds?
A one-pass run treats the request as a single bounded job. The agent reads the goal, calls any tools it needs, and returns a result. It does not reopen the goal after the first result unless the person sends a new request.
Working in rounds treats the request as an objective. The agent takes a step, inspects the new state, and then asks whether the objective is met, blocked or still open. If it is still open and the person's limits allow it, the agent takes another step. That loop can produce a better result, but it can also consume more time, spend and attention than a one-pass run.
What should 'done', 'blocked' and 'still open' mean?
These states are useful because they separate completion from progress. A run is not done simply because the agent produced text or used a tool. The outcome should be compared with the objective and the preconditions the person set.
Blocked is not failure. It means the next step depends on something the agent does not have: access, a file, a decision or a payment. Still open means there is a legitimate next step, but the work is not finished. Clear definitions stop the agent from declaring victory too early or continuing when it should ask.
| State | Meaning | Next action |
|---|---|---|
| Done | The objective or approved sub-task is complete, and no open dependency remains. | Record the outcome and stop unless the person opens a new round. |
| Blocked | The next step needs input, access, payment, a missing file or a human decision. | Pause and ask; do not improvise around the missing item. |
| Still open | Work has progressed, but the objective is not yet met and a clear next step remains. | Continue only if round, time and spend limits still allow it. |
| Precondition not met | A required condition has failed, so further attempts would be wasted. | Stop and report the failed condition before any retry. |
Which limits should a person set before an agent starts?
The limits should be attached to the objective, not discovered halfway through a run. Once the agent is in a loop, it becomes harder to judge whether another round is worth it, so the caps need to exist before the work begins.
These limits do not prevent ambitious work. They make the agent return to the person at a defined boundary and explain what has changed. The person can then open another round, widen a cap or stop.
- Round cap
- The maximum number of work cycles the agent may take before it must report back.
- Time budget
- A wall-clock or active-run limit after which the agent must stop, even if progress seems possible.
- Spend cap
- A cost ceiling that prevents the agent from consuming further paid actions or model calls.
- Approval checkpoint
- A required human confirmation before a stage such as sending, publishing, signing or payment.
Why should the stop decision belong to the person rather than the model?
A model can judge whether a task looks complete by its own output, but it cannot judge your budget, your risk appetite, or the cost of a marginal step. It may propose another round because a further step seems possible, not because it is worthwhile.
The person should hold the stop because the person owns the consequences. The agent should report the state, the reason it stopped and what it would do next, but the decision to extend past a cap should be explicit.
What should you expect to see when a run ends?
A clean end is not just a final answer. It should include the outcome state, the artefacts produced or changed, any open items, and the trigger that stopped the run. That way you can decide whether to continue without reconstructing the agent's path.
Good output includes enough context to make the next decision quickly. If the stop was a round, time or spend cap, the agent should say what remains and what it would do in the next round. If the stop was completion, it should say what changed and where.
- the outcome state: done, blocked, still open or precondition not met
- the concrete artefacts created or changed
- the trigger that stopped the run
- any open item that would need input or access
- the next action the agent proposes
Common questions
Can I let an agent run without limits if the task is important?
Important tasks benefit from limits more than routine ones, because the consequences of a wrong step are larger. A cap does not stop useful work; it creates a checkpoint where you review progress. You can always extend the run after that review. Without limits, you may return to find the agent has spent heavily on a path you would not have approved.
Does a single-pass agent always finish faster than a multi-round agent?
Not necessarily. A one-pass run returns sooner, but it may miss a dependency or leave work incomplete. A multi-round run can take longer, yet it may finish the objective properly. The relevant question is not speed alone, but which style fits the task: a lookup suits one pass, while a multi-step objective may need rounds under explicit caps.
What should count as blocked rather than still open?
Blocked means the next step cannot proceed without something the agent does not have: access, a file, a human decision or payment. Still open means a clear next step exists and the agent could take it within the current limits. If the agent can proceed but the limits are exhausted, that is a stop, not a block.
Who should approve continuing past a limit?
The person who set the limit should approve any extension. The agent may recommend continuing, but it should not silently widen its own caps. A short report of what changed, what remains and what the next step is gives the person enough to decide whether another round is justified.
For a practical view of how run limits and statuses are handled in one agent system, see the help centre or compare plans.