Two settings shape how the AI behaves while it works on a workflow step:
Error handling policy — what the AI does when a tool call fails: stop, use its own judgment, or keep going.
When the AI saves results — whether the AI writes results to the work item as it gets them, or once, just before it finishes or pauses the step.
Both work the same way. You set a thunk-wide default, and any step can override it. They apply every time the step runs, including when scheduled work or an incoming message triggers it.
Error handling policy
Choice | What the AI does when a tool call fails | When to use it |
Conservative | Stops the step with an error on the first tool failure, instead of working around it. It can also stop early when it detects a logical error of its own. | A wrong result is worse than no result, for example a step that pays, files, or sends something. You would rather see the error and fix the cause. |
Balanced (default) | Uses its own judgment. It gets no special instruction, and a failed tool call does not stop the step on its own. | Most steps. |
Proactive | Is told to keep going after a tool failure and try to finish the step. | Steps where a partial answer is still useful, and one failed lookup should not hold up the work item. |
When a Conservative step stops, it ends with an error you can review. See Troubleshooting errors for how to read it, and How an AI agent ends a run for the other ways a run can end.
When the AI saves results
Choice | What the AI does | Trade-off |
As it works (default) | May save results to the work item whenever it has them, for example after each search. | Work already saved survives if the step times out or fails. It takes more AI turns, and people looking at the work item can see values that are not final yet. |
At the end | Keeps its results in the step chat and saves them once, just before it finishes or pauses the step. | Fewer AI turns, and no half-finished values on the work item. But results not yet saved are lost if the step times out or fails. |
Choose At the end for short steps where only the final result matters. Keep As it works for long or expensive steps, where partial progress is worth keeping.
This setting guides the AI. It is not a hard rule: under At the end, the AI can still save something early, and under As it works, it can still save everything at the end.
Using both settings together
At the end and Conservative are a risky pair. A Conservative step stops on the first tool failure, and under At the end the AI has usually not saved its results yet, so they are lost. If the step then runs again, it has to redo that work. If a step needs Conservative error handling and runs for a long time, keep As it works.
Set the thunk-wide default
Open the thunk's Settings.
Select the Reliability Guardian section.
Choose an Error handling policy and a setting for When the AI saves results.
Every step set to Inherit uses these defaults. Reliability Guardian settings depend on your plan.
Override the default on a single step
Open the step in the workflow plan.
Expand its AI Settings.
Under Error handling policy or When the AI saves results, choose a setting other than Inherit.
Inherit is the default for every step, and uses the thunk-wide setting. Only the thunk's owner and admins can change these settings. Changes show up in the step's version history.
For the other controls that make a step more reliable, such as its Directions, properties, and tools, see AI reliability mechanisms.