How the agent runs changes the product
A synchronous API agent, a background worker, a scheduled monitor, and an event-driven webhook agent have different failure modes. Choose the deployment shape before pretending the prompt solved architecture.
Long-running agents usually belong behind queues or job runners. Users get status, cancellation, and partial results instead of a hanging request.
Graceful failure is a feature
When an agent fails, it should preserve state, explain what was completed, identify what failed, suggest next action, and avoid pretending success.
The best failure mode is often a handoff: to a human, a simpler workflow, or a saved draft. Silent failure is betrayal. Loud failure with state is recoverable.