These steps map to tools in Capabilities. You don’t call them by hand — describe the outcome you want and let the client chain the tool calls. Example prompts are shown throughout.
1. Branch
Never edit a live agent directly. Create a branch so your changes are isolated until you’re ready to merge.
“Create a branch of the Acme support agent called greeting-update.”
2. Build
Edit the agent on the branch — behavior rules, knowledge base topics, connectors, and variants for A/B experiments.“On that branch, update the greeting and add a knowledge base topic for refund policy.”
3. Test
Use a debug chat session to exercise the branch before it goes anywhere. Inspect the resulting conversation data to confirm the agent behaves as intended.“Start a debug chat against this branch and ask it about refunds. Show me the transcript.”
4. Merge
Once the branch checks out, merge it back into the agent’s main line.
“Merge greeting-update back into the agent.”
5. Deploy to staging
Publish the current draft to a non-production environment first. Test again against staging with real routing.“Publish the current draft to staging.”
6. Promote to production
When staging looks good, promote the deployment to production. Nothing reaches live callers until this step.“Promote the staging deployment to production.”
7. Roll back if needed
Every deployment is versioned. If something regresses, roll back to a previous deployment immediately.“Roll production back to the previous deployment.”