Skip to main content
Athena combines agent building with tools for choosing a model, routing calls, managing batches, and reviewing what happened. This page points to the guides for each workflow and calls out settings that affect a live conversation. Feature visibility and provider support can vary by organization and call path; verify the choices available to your account before rollout.

​
Build and release an agent

Create a workflow, connect its conversation steps, add the model and tools it needs, then validate and publish the definition you intend to use. Dashboard edits can remain a draft while published call triggers continue using the last published version. The /test/ phone routes select a draft but can still place real calls.

​
Configure voice behavior

The workflow Settings page includes voice behavior controls. Changes saved to a draft apply to draft tests; publish the workflow to apply those settings to published calls.
  • Caller Memory can give a realtime voice agent the caller’s last three conversations from the previous 90 days. Phone calls use earlier calls from the same number; browser tests use earlier tests of that agent. Use the Remember Previous Conversations toggle to turn it on or off. Review whether this behavior fits your caller experience and data-handling practices.
  • Ambient Noise adds background audio. You can use the default office ambience or upload a custom audio file up to 10 MB and adjust the volume from 0 to 1. The run path you select must support this setting.
  • Phone Menus (Keypad) lets a phone agent listen to an automated menu, press a key, and speak after a person or department answers. Wait Quietly on Hold suppresses agent check-ins while the other side plays hold music, tones, or recorded messages; the agent speaks again when a person returns.
  • Call Management includes a maximum call duration and maximum user idle timeout. Set limits that match the task and verify them in a controlled test.
Phone-menu behavior and Retry Review’s unusual-menu hold are related but separate: the first helps an agent navigate a menu during a call; the second decides whether a failed call should be tried again. See Retry Review.

​
Choose models and fallback rules

A provider connection holds account settings; a named model configuration chooses services and model settings; the agent can inherit a default or save a workflow-specific override; supported call fields can add request-level overrides. Configure fallback LLMs for a slow response or a pre-output request error when appropriate. A fallback can send conversation data to another configured provider and add cost; it does not switch the realtime audio model. See Providers, models, and voices for the configuration layers, voice/language selection, and fallback behavior.

​
Manage campaign traffic

Campaigns can select one agent version or divide traffic among published/archived versions. A variant can follow the latest published version or stay pinned to a numbered release. Schedule calls in a chosen time zone, set rate and concurrency within the organization’s limits, and configure campaign retries for selected outcomes. See Run campaigns for CSV setup, traffic splits, calling windows, limits, and the difference between campaign retries and Retry Review.

​
Review failures and call history

Retry Review groups recent failed calls by likely cause and current next step. It can offer an immediate rerun, a waiting-line action for a multi-call selection, or enqueue the work in its campaign. Repeated phone-menu failures can be held; a cause that needs a correction is not fixed by dialing again. Reruns are real calls, and eligibility is checked again when requested. Open Retry Review for outcome groups, attempt chains, blocked_reason, and the rules shared by the dashboard and API. For completed calls, the run page combines call details with Athena Live activity, transcript and media links when available, and separate caller/agent recordings when those tracks exist. Reports can summarize one workflow, a campaign, or daily results. See Calls, runs, and results and Troubleshooting for a practical review path.

​
Add text interaction

Use the generated embed widget with an origin-restricted embed token, or use the authenticated text-chat API from your backend. The authenticated API supports session recovery, revision checks, POST-based SSE, rewind, and end. Public embed tokens are designed for visitor sessions and do not expose private tool results.