Loops, failures & approvals
By the end of this page you will know how a turn recovers — when work goes back for another pass, when a step fails, and when an approval has nobody to answer it.
Prerequisites
- A Planner / Executor system with at least one member (Your team).
- Admin or Developer role.
4. Redoing work
A reviewing agent can send work back to the agent that produced it — a fact-checker returning a draft to the writer, for example.
| Control | What it does | Default |
|---|---|---|
| Let agents send work back | Allows the planner to propose bounded retry loops. | On |
| Most attempts per loop | How many times one loop may repeat. After this the turn finishes with what it has, and says so. | 3 |
Each attempt is a full agent run, and it draws on the same Total agent runs per turn budget as everything else. A loop that repeats three times over two agents spends six runs of that budget.
The loop always ends. When the attempts run out, the turn completes with the best result it has rather than failing.
5. When a step fails
If a step fails outright, the planner can write a new plan for what is left, using whatever already succeeded.
| Control | What it does | Default |
|---|---|---|
| Re-plan after a failure | Lets the planner replan around a failed step. | Off |
| Most re-plans per turn | How many times it may do that. | 1 |
Completed work is never redone. A re-plan covers only what is left.
It is off by default because it doubles the planning cost of a failing turn. Whether you prefer a fast honest failure or an expensive recovery depends on your workload, so it is not something a default should decide.
All re-plans share the turn’s agent-run budget and its time limit, so turning this on does not raise your ceiling — it spends the same ceiling differently.
Approvals when nobody is there
Some tools pause and ask a person before they run. In a chat, that is fine — someone is there. On a scheduled, triggered or API turn, nobody is.
The When nobody is available to approve setting, in the Advanced section, decides what happens then:
| Option | What it does | Use it when |
|---|---|---|
| Wait for approval | The step parks and waits for someone to answer. | The work can sit until a person is available. This is the default. |
| Treat as rejected | The step proceeds as though a person declined. | Declining is a safe, meaningful outcome — skipping an optional notification, say. |
| Fail the step | The step fails, and the turn handles it as any other failure. | Silence should be treated as a problem, not an answer. |
This setting only applies to scheduled, triggered and API turns. Chat turns always ask, because someone is there to answer. Changing it will not affect anything you see in the preview.
How these interact with the budget
All three features spend the same Total agent runs per turn budget:
- A loop attempt is an agent run.
- A re-plan’s new steps are agent runs.
- A parked approval holds its step, and the turn’s time limit keeps running.
The turn budget bar in the configuration panel shows the worst case with your current settings. If you turn on loops and re-planning and leave the budget at its default, check that bar — it will tell you whether a bad turn can finish.
Save your changes
Click Save in the header, then test in the preview.
Common issues
- A loop never stops — It does. Most attempts per loop bounds it, and the turn completes with what it has.
- The turn stopped early and said so — It hit Total agent runs per turn. Raise the budget, or lower the attempts and re-plans feeding it.
- Re-planning had no effect — It is off by default. Turn on Re-plan after a failure.
- A scheduled run is stuck waiting — Wait for approval is the default and nobody answered. Choose Treat as rejected or Fail the step for unattended work.
- The approval setting seems to do nothing in the preview — Chat turns always ask. Test it on a scheduled or API turn.