Before you begin
Sign in to Athena and select the organization where the agent should live. Check that the organization has an available model configuration. If you plan to test by phone, you also need an outbound-ready telephony configuration and a phone number you control. For a browser voice test, allow microphone access. Use synthetic names, account numbers, and case details while building. A test can create a saved run, consume model capacity, call a real phone number, or send a real request through a tool.Create and prepare the workflow
1
Create an agent
Choose Voice Agents, create an agent, and give it a name that makes its purpose clear to your team.
2
Describe the job
Write instructions for the agent’s role, what information it may collect, the actions it may take, and when it should end the conversation. Avoid putting credentials or other secrets in instructions.
3
Build the conversation path
Add and connect the workflow steps that implement the task. Cover the expected answer, a caller who cannot provide it, and a request that needs a human or another route.
4
Select dependencies
Choose an available model configuration. Add only the tools and knowledge the workflow needs; test each external endpoint separately with synthetic inputs.
5
Validate and test
Resolve validation errors, then exercise the workflow with representative inputs. Review the saved run, tool results, and any external receiver logs.
6
Publish deliberately
Publish the reviewed definition when it is ready for the published trigger endpoints. Record a version name and change note if your team uses release metadata.
Test the right thing
The dashboard’s test controls and the public phone endpoints answer different questions. A browser conversation can help you check prompt flow and tool use, but it does not prove phone routing. A phone endpoint containing/test/ selects the latest draft (or the published definition when no draft exists) and can still dial a real number. It is not a dry run.
For a phone test, use a controlled destination and verify each stage: request accepted, provider accepted, handset rang, two-way audio worked, the call ended as expected, and the run shows the intended result. An HTTP success alone proves only that the initiation request was accepted.