Configure alternatives deliberately
When creating an agent, you can provide fallback models for cases where the primary model fails. Select alternatives that can handle the same input types and output requirements as the task. If the run depends on a particular capability, such as image input or a strict response schema, check that the fallback supports it before relying on the configuration.
Fallbacks are most useful when a task can tolerate a different model producing the answer. They may vary in style, interpretation, or detail, so do not assume the replacement will behave exactly like the primary choice. For high-impact work, keep an approval or review step after generation even when a fallback lets the run complete.
Check which model completed the work
After a run, inspect its result and available run details to see whether it finished as expected. If the output appears incomplete or differs from the requested format, validate it against the task requirements instead of treating successful completion as proof of quality.
If no configured model can handle the input, the run may still fail. In that case, check the model configuration and the task's input and output needs before retrying. Recording which model was intended and what result was produced makes it easier to diagnose an intermittent issue and decide whether a retry is appropriate.