Skip to main content
Support assistant tutorial You now have the building blocks: an Agent, read tools, conversation state, retrieval, approval, streaming, and an HTTP boundary. Deploying them requires decisions about your application’s users, data, and operations.

Define the contract

Build a small evaluation set

Include the prompts you tried in this tutorial: a known order, an unknown order, a follow-up in the same session, a follow-up in a different session, a supported policy question, an unsupported question, and both approval outcomes. Assert the underlying effect and retrieved evidence where possible. For the approval case, the refund counter is a better contract than a substring in the model’s answer. For a real provider, pin the model configuration and record it with the evaluation results. Evaluation covers suites and scorers. Troubleshooting helps distinguish an execution failure from an answer-quality problem.

Decide what needs recovery

Conversation persistence preserves history. Queue redelivery schedules another attempt. A checkpoint captures state. None of these alone knows whether a payment provider already accepted a request. Choose the recovery contract before enabling automatic retries for writes.

Continue by task

Authenticate and authorize

Bind execution to verified identity and resource ownership.

Plan recovery

Choose history, queues, or durable tasks for the actual failure.

Operate the application

Set budgets, collect traces, handle shutdown, and investigate errors.

Choose more structure

Add a harness, workflow, or team when a concrete requirement calls for it.