Agent Mode Troubleshooting

The run stopped before another call

Check the run’s error and Costs and Limits. Insufficient wallet funds, a reservation larger than the remaining budget, a missing model price, or a planner context over 32 KB can stop a run before dispatch. Add funds, choose a cheaper available model, or shorten the task as appropriate.

Settlement pending or provider outcome unknown

Keep the run ID. View its detail and Usage before starting another run. Known usage can settle later; an unknown provider outcome retains its reservation until reconciliation. Contact support if it remains pending. These states do not prove that generation failed or that no charge occurred.

The planner returned invalid output

The provider may have completed a billable call without returning a usable tool decision or final answer. Its usage is still recorded. Review the error, model’s tool support, and recorded cost before choosing a new run.

A tool failed

Read the tool’s observation in the timeline. A project can restrict tools, artifacts must be accessible to your account, and external pages can be unavailable. Repeated tool failures stop the loop. A final synthesis is attempted only if budget and funds permit it.

The cost is small or the run is old

New runs display sub-cent settled charges. A pending call is not included until it settles. Historical runs before the September 5, 2026 accounting correction keep their original recorded values and were not retroactively charged.

Stop or disconnect

Stop is cooperative: an in-flight provider call can still finish and settle. Closing the browser does not cancel or refund provider work. Preserve the run ID and check run status before retrying.