An agent demo shows that a task can be completed. A production rollout needs to show what happens when inputs are incomplete, a tool fails, or the result needs human review. Start with a narrow task and a clear definition of success.
Define the boundary of the task
Use a predictable workflow where the steps are known. Introduce model-directed decisions where they add value, and make the agent’s tools easy to understand. Anthropic’s engineering guidance recommends starting with simple patterns and increasing complexity when the task requires it.
Make failure visible
A useful tool contract distinguishes a completed action from an error or an uncertain outcome. For example, a payment request that times out may still have reached the provider. Record the request identifier and check its status before retrying. The agent should have a way to stop and ask for help.
Limit what the agent can do
Define allowed tools, access permissions, and execution limits before connecting live systems. A practical review should cover who can approve sensitive actions, which writes can be reversed, and who owns an exception after escalation.
Test the whole workflow
Build a test set around the actual task: complete inputs, missing fields, conflicting information, unavailable tools, and requests outside the agreed scope. Assess whether the action and final record are correct. Repeat these checks when prompts, models, or integrations change.
Before the first rollout
- Name the owner of failed or escalated tasks.
- Set a limit on steps, retries, and execution time.
- Keep a record of inputs, actions, and results with appropriate access controls.
- Agree on a rollback or disable procedure.
- Start with a limited release and review the exceptions.
References
Primary documentation used for this article. Product features and pricing can change.
