A run stops after an external action but before recording its response. Resuming from a text conversation can misrepresent what actually happened.
Record meaningful state
Track task identity, completed steps and pending operations. The service result, rather than the agent's wording, determines whether an action succeeded.
Reconcile before retry
Check available evidence from the target system. Some operations support safe repetition; others require manual review. Document that distinction.
Simulate interruption
Drop the response after a permitted action in a test environment. Ask the supplier to show recovery without duplicate effects. Include ambiguous outcomes in acceptance. Resume behavior should be implemented and verified in the application, not left to an instruction asking the model to remember its progress.
Related reading : Agent costs: compare successful jobs rather than individual model calls. Discuss your project scope with Sky Vaults.