Test an agent

Hear your agent in the browser

Open Agents, select your agent, configure its supported STT, LLM, and TTS components, then start a browser test. Allow microphone access and end the session when you finish. Each session is capped at 180 seconds.

Authenticated browser tests use welcome browser credit or purchased funds at the selected components' metered prices. You do not need a phone number for a browser test. This does not qualify the carrier route or prove that a real phone call will work.

If the AI agent does not start speaking

If the requested opening fails before the worker observes agent audio, the browser test ends and explains that the AI agent did not start speaking. Start another test to retry with the same selected configuration. Each retry creates a new call and passes the usual balance and admission checks; it does not retry automatically.

If a connection ends before its outcome can be confirmed, the panel says so. This message does not mean your configuration is wrong or confirm a refund. Microphone resources are closed before you retry. Ending the test yourself keeps the normal end-of-test behavior.

A confirmed zero-output startup failure can release an unused hold when that decision is processed before native usage is accepted. If usage arrived first, normal settlement applies. Missing or incomplete evidence keeps the pending-usage hold until usage arrives or the 24-hour recovery returns it; silence alone is not a billing decision. See Billing.

For integrations, GET /v1/calls/{call_id} may expose browser_outcome while the call is still active: browser_startup_failed, worker_ended, or an empty string when no outcome is known. worker_ended confirms teardown, not billing success. The transport still completes through its normal lifecycle. A failed opening has ended_reason: browser_startup_failed after completion. These fields do not certify a refund.

Chat with your agent in text

The same test panel has a Chat mode. For an AI agent built from instructions, an opening line, tools and knowledge — one without a conversation flow — Chat runs the agent as a browser test would, minus audio: the same resolved configuration, the same language, the same tools and attached knowledge, answered by the agent's configured model. Choose the draft snapshot to talk to the unsaved edits on screen or published to talk to what callers reach; switching between them starts a new conversation.

The opening line arrives without a model call when one is configured, or is generated from the instructions when none is. Every tool the agent calls during a turn is shown in the transcript:

• search_knowledge_base runs the real retrieval against the attached bases.
• end_call ends the conversation, as it ends a call.
• SMS, transfers, webhook and integration tools are held — the agent is told the action did not happen and the transcript marks the call as held — so a text test never sends a message, dials anyone or reaches a customer system. Use a browser or phone test to exercise those for real.

Text chats are not metered against your funds and create no call record; only the model runs, and the API reports the model calls and tokens each turn used. A conversation is bounded to 2,000 characters a message, 40 lines and 24,000 characters in total, and ends when it reaches that length. The same test is available to integrations as POST /v1/agents/{agent_id}/chat-turns — the client keeps the transcript and sends it with each turn, and may send test_values for the agent's call variables, checked and filled in exactly as a call would admit them; see the API reference.

An AI agent that has a conversation flow keeps its flow-driven Chat, described under Agent workflows.

Test a real phone call

For a PSTN test, use a normally configured number and carrier route, purchased funds, and the usual authorization and policy checks. Browser credit cannot fund the phone call. See Wire an agent and Policy and outbound.

Shared-number PIN trials are retired

New customer-configured live shared-number trials are not available. A PIN from create_trial_call is not a number you can dial to reach your agent.

The tool and POST /v1/trial-calls remain for offline synthetic fixtures and require synthetic_only: true. They do not create real audio or prove a working phone route. Live routing/completion through the old trial path is rejected. Existing trial records remain readable. This change does not remove the separately bounded anonymous homepage demo.

Try the homepage demo

The anonymous homepage demo uses shared demo capacity, separate from your workspace's browser tests and ordinary phone calls. If it is busy or temporarily paused for maintenance, wait and try again. Adding wallet funds does not open a demo slot, and starting a phone call is not a substitute for retrying the demo.

Maintenance can pause new demo sessions while existing sessions finish. Demo availability does not establish how many simultaneous phone calls your workspace can make.

Review the result

Use Calls or the call-evidence tools to inspect status and available recordings and transcripts. A failed call or a synthetic fixture need not have real audio evidence. See Read call evidence.