Agent demos usually fail in production for ordinary software reasons: unclear permissions, weak error handling, missing measurements, and no safe way to stop. The model is only one component of the system.
1. Give every tool a narrow contract
Define accepted inputs, returned outputs, timeouts, and failure states. Reject unexpected fields before a tool reaches a database, browser, or external API.
2. Separate suggestions from actions
An agent may draft an email, refund, deployment, or database change. A person or policy engine should approve consequential actions before they leave the system.
3. Test the failures you expect
- Malformed or adversarial tool output
- Rate limits, timeouts, and partial responses
- Prompt injection inside retrieved content
- Loops that consume the entire token or cost budget
4. Measure outcomes, not confidence
Track task success, groundedness, latency, cost, human corrections, and unsafe action attempts. A confident answer is not evidence that the task was completed correctly.
The production checklist
Ship with least-privilege tools, explicit stop conditions, durable traces, representative evaluations, privacy rules, human approval, budget limits, and a rollback plan. If one is missing, the agent is still a prototype.