ElevenLabs Agents (ElevenAgents)

Production guidance for building voice (and text) agents on the ElevenLabs Agents platform. The product was renamed from "Conversational AI" to "ElevenLabs Agents" / "ElevenAgents" during 2026; documentation URLs use both /docs/eleven-agents/... and legacy /docs/agents-platform/... paths, and the two mostly mirror each other. All model IDs, prices, limits, and endpoints below are volatile and were verified on 2026-07-10 against elevenlabs.io/docs — re-verify before quoting them to a user, because ElevenLabs adds LLMs and revises plans frequently.

When this skill applies

Use this skill when the work is a two-way spoken (or typed) conversation driven by the ElevenLabs Agents runtime: phone bots, website voice widgets, in-app voice assistants, WhatsApp/chat agents, outbound call campaigns, or a custom SDK integration. It also applies to configuring the agent's brain (prompt, LLM, tools, knowledge), its voice, its channel, and its evaluation.

Do not use it for one-shot text-to-speech, audiobook narration, dubbing, sound effects, or music — those are separate ElevenLabs products with their own APIs. If the user only wants a .mp3 from text, they want the TTS API, not an agent.

The pipeline you are configuring

[Documented] An ElevenAgents conversation is an orchestrated loop over four components (elevenlabs.io/docs/eleven-agents/overview, verified 2026-07-10):

1. Speech-to-Text (ASR) — a fine-tuned recognition model that turns caller audio into text. ElevenLabs' custom ASR is quoted at sub-100ms versus ~300ms+ for a generic Whisper deployment.
2. Turn-taking / interruption model — a proprietary model, layered on a Voice Activity Detector (VAD), that decides when the user has actually finished a turn (vs. a mid-sentence pause) and when a barge-in should stop the agent.
3. LLM — the reasoning brain. You pick from hosted models (Google/OpenAI/Anthropic/ElevenLabs-hosted open weights) or bring a custom endpoint.
4. Text-to-Speech (TTS) — a low-latency voice model rendering the reply, drawn from 5,000+ voices across 70+ languages (agent voice picker exposes ~31 languages; Flash/Turbo TTS models cover ~32).