Sandbox gives your team a place to develop agent behavior before exposing it to customers. Use it to prepare instructions, knowledge, tools, and procedures, then test the saved draft against the requests you expect to receive. A release is the reviewed version you publish to Production. Keeping the draft and published behavior distinct makes it easier to improve one workflow without treating unfinished experiments as customer-ready.

​
Define expected outcomes before testing

Write representative questions and the result you expect for each one. Include the answer source, any required action, and whether the request should go to a person. For a returns workflow, test an eligible request, a request outside the policy window, and a case where the policy is missing an exception. For a connected lookup, add a missing record and a service failure.

​
Test in Playground

Use Playground and personas to explore realistic variations in how customers ask for help. Check the final answer and any tool behavior, rather than judging only the tone of the response. When a test fails, identify whether the cause is knowledge, instructions, routing, permissions, or the connected system. Revise the relevant part and rerun the scenario. Keep a record of the failure and the corrected result so that later changes can be assessed against the same expectation.

​
Save and validate the draft

Save the Sandbox changes and run Validate. Fix blocking configuration errors before publishing. Passing validation establishes that the configuration satisfies the available checks; it does not prove that every customer answer will be correct. A manager should still review the important outcomes, especially identity boundaries, sensitive actions, and handoffs.

​
Publish the reviewed changes

Use Set Live to create the Production release. The first release publishes the full agent. Later releases can select saved components, allowing reviewed changes to move forward while unrelated draft work remains unpublished. Confirm the selection before release. Production versions provide a record of what was published; do not assume a particular rollback workflow without checking the controls available in your release.

​
Test the deployed experience

After publishing, try the chosen channel with fictional records. Check the actual conversation, an appropriate handoff, and a failed lookup where relevant. A correct Playground result does not replace this channel check. Continue monitoring the conversations that follow. Use them to refine the next draft, and repeat the tests whenever a policy, integration, or important instruction changes. The 14-day evaluation plan provides a practical sequence for your first trial.