Skip to main content
Your first Chatropic agent should have a clear job. Rather than connecting every system at once, choose a support request your team understands well and work through the steps needed to handle it reliably. This guided tour follows that process from preparation to review. It describes the Builder workflow; it is not a recording of a customer deployment or a promise about how long your implementation will take.

​
1. Define the agent’s responsibility

Create a Super Agent with a business objective and clear instructions. Describe the requests it should handle, what it should avoid, and what a successful outcome looks like. For a fictional retailer, the initial objective might be to answer delivery-policy questions and explain when a request needs the support team. Specialist subagents can be added for narrower responsibilities as the workflow grows.

​
2. Add approved knowledge

Prepare the material the agent needs: current policies, help articles, product guidance, or other approved content. Organize it in Knowledge and check the indexing state after uploading a source. Test a few ordinary questions before adding more material. If the answer is wrong, inspect whether the source is missing, outdated, contradictory, or difficult to interpret. More content is not always the right fix.

​
3. Connect the information that must stay live

Use a data connector or permitted tool when the answer depends on a current customer record. A delivery policy belongs in knowledge; an individual order’s dispatch status belongs in the order system. Engineering support may be needed for authentication, customer identity, or custom operations. Begin with the minimum access needed for the selected workflow and separate lookups from actions that change records.

​
4. Test before customers use it

Use Playground in Sandbox to try realistic questions and different personas. Include unclear requests, missing records, failed tools, and questions that should go to a person. Keep expected outcomes beside each test. “Sounds helpful” is weaker evidence than “returned the correct status and escalated the disputed delivery.” Save and revise the draft until the selected workflow meets your agreed threshold.

​
5. Validate and publish the reviewed changes

Run Validate on the saved draft and resolve blocking errors. Use Set Live to publish a Production release. The first release publishes the full agent; later releases can select saved components. Validation checks configuration. You still need to test the actual deployment channel and customer experience after release.

​
6. Review conversations and improve

Inspect conversations, tool evidence, and human handoffs. Use what you learn to improve the next Sandbox draft. Open Reports to review AI performance, Topics explorer, and Content gaps. Use a detected gap to draft an article, then test and publish the approved answer. Start with the trial preparation guide to gather the content, people, and test cases you will need.