How to test an ElevenLabs voice agent

ElevenLabs Agents already ships most of a first test pass: three kinds of automated test, a scoring system that grades every finished conversation, and webhooks that send the results to your own systems. A handful of phone calls to the number callers will really dial covers the rest.

Key facts

• ElevenLabs offers three kinds of agent test: simulation, next reply and tool call, per its Agent testing page.
• Evaluation criteria return success, failure or unknown, per the success evaluation page.
• The default maximum call duration is 600 seconds, per the conversation flow page.
• Transfer to a number works on phone calls only, not in the chat widget, per the transfer page.
• Agents made through the API or SDK do not get the end call tool unless you add it, per the end call page.

What can ElevenLabs test for you?

The Agent testing page describes three kinds of test.

• Simulation: a simulated caller holds a full multi-turn conversation with the agent, and the test checks whether the conversation reaches the outcome you defined. You describe the scenario and a success condition, and you set a turn limit of up to 50.
• Next reply: only the agent's next message is scored, against success criteria and example good and bad answers that you supply. An LLM judges it.
• Tool call: checks that the agent calls a named tool with the right parameters. Parameters can be matched exactly, by regular expression, or by an LLM judge.

Simulations can mock tools, so a booking test does not book anything. You can mock none, all or selected tools, though the page says system tools and workflow tools are never mocked.

Open a call in the conversation history and choose "Create test from this conversation." A real failure becomes a regression test in a few clicks.

Run tests from the dashboard, from the command line with elevenlabs agents test <agent_id>, or through the API. A language model plays the caller, so one pass proves little. The API's run-tests call takes a repeat_count of 2 to 20, and the page describes badges that go green at 100% and amber at 80% or more, with failures grouped into buckets. A flaky pass is itself a finding.

How are finished conversations graded?

Two features score each real conversation once it ends, and both are set up in the agent's Analysis tab.

Evaluation criteria are yes-or-no questions written as prompts, such as "the caller was not left upset". Each returns success, failure or unknown, with a written rationale. The page says unknown often means the conversation was incomplete. Results show in the conversation history and in post-call webhooks.

Data collection pulls structured fields out of the transcript. Each field has an identifier, a type (string, boolean, integer or number) and a description. For testing, this is how you check that the agent captured the right date, name or reference number from a call.

A language model reads the transcript to grade both, so neither listens. They will not notice a clipped word or a long pause, and their verdicts should be spot-checked against the recording.

Where do ElevenLabs test results go?

Open the conversation history for transcripts and the analysis results. To pull the same data into your own tools, use the post-call webhooks. There are three types: post_call_transcription (transcript, analysis and metadata), post_call_audio (the full recording as base64 MP3), and call_initiation_failure for calls that failed to start.

Verify the ElevenLabs-Signature header before you trust a payload, and return a 200 status. The page says a webhook can be switched off automatically after repeated failures, so a dead test endpoint can stop results arriving without any visible error. Audio retries resend the whole payload, so deduplicate on conversation_id.

The Get conversation endpoint returns the transcript, with a time in the call for each turn, plus the analysis. Its audio fields are only flags that say whether a recording exists, so use the audio webhook or the dashboard for the sound itself.

Which settings need a real call?

Simulations run on text and transcripts, so these settings only show their real behaviour when someone speaks to the agent on a phone.

• Turn taking: the conversation flow page lists a turn eagerness setting (eager, normal or patient), a "take turn after silence" timeout, and a soft timeout that makes the agent say something when the caller goes quiet. Try a caller who reads a long reference number with pauses. More background: turn-taking and latency.
• Interruptions: the same page says interruptions are enabled through the client events in the Advanced tab. Cut in early, mid-sentence and near the end of a long answer. See interruption handling.
• Maximum duration: the same page gives a default of 600 seconds, which a slow caller can reach on a long call. Listen to what the agent does when time runs out.
• Transfer to a number: the transfer page says the tool works on phone calls only, and not in the chat widget. It offers conference, blind and SIP REFER types, and some depend on the native Twilio integration. A text simulation cannot prove the call really reaches a person, so ring the destination yourself.
• Transfer to another agent: the agent transfer page says the transcript carries across, but most other settings, such as prompt, voice and tools, come from the receiving agent. Check that the second agent does not ask for what the caller already gave.
• End call: the end call page says agents made through the API or SDK do not get the tool unless you add it. A missing tool means the agent never hangs up on its own.
• Language: the language detection page says the tool is not enabled automatically and that background speech or an unusual accent can trigger an unintended switch. Test with accented callers and with a noisy room. See phone audio, accents and noise.

Which calls should you run?

Run each of these several times, since the same call can go differently on a second try. RingTru's simulated callers can ring the number for you with accents, noise and interruptions.

• A caller who completes the main task with no trouble, as a baseline.
• A caller who changes their mind halfway, such as moving a booking from Tuesday to Thursday.
• A caller who talks over the agent during a long answer, and one who says only "mm-hm".
• A caller who pauses while reading out an account or reference number.
• A caller who asks for a human straight away. Confirm that a phone actually rings and that the caller hears a sensible message while they wait.
• A caller who asks for something the agent should refuse or cannot do.
• A caller who goes silent for a long time.
• A caller who speaks another supported language, or switches mid-call.
• A caller who says goodbye. Check the call ends cleanly and quickly.

After each call, read the transcript, listen to the recording, and compare what the evaluation criteria and collected data say with what you heard. Where they disagree, fix the criterion before fixing the agent.

Platform features change often. Everything here was checked against the linked pages on 27 October 2026. For why text simulations are not enough, see text vs audio testing. To compare with other platforms, read how to test a Vapi voice agent and how to test a Retell AI voice agent. For the general approach, see voice agent testing and the guides by platform.