SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillselevenlabs-agents

elevenlabs-agents

Agents & models

View repository

Updated 10-15 20:09

About this skill

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).

Your job is to choose and tune each stage so the whole loop feels like a person. [Heuristic] Latency and quality trade against each other at every stage; optimize the loop as a whole, not one component in isolation.

Latency: the budget that decides the feel

[Documented] ElevenLabs advertises first-turn latency under ~500ms for the ElevenAgents 2.0 stack (elevenlabs.io blog, 2026). Approximate per-stage contributions from ElevenLabs' own latency write-up (elevenlabs.io/blog/how-do-you-optimize-latency-for-conversational-ai, verified 2026-07-10):

Stage Typical contribution
ASR (ElevenLabs custom) <100 ms
Turn-taking / VAD silence detection tunable; adds the "end-of-turn" wait
LLM first-token Gemini Flash <350 ms; GPT-4-class / Claude Sonnet 700–1000 ms
TTS (Flash v2.5) ~75 ms model time, ~135 ms end-to-end
Network (same-region) ~200 ms+
Telephony (global) ~500 ms+

[Independent benchmark, disclosed method] Third-party monitoring (Cekura, Deepgram, 2026) reports ElevenLabs among the fastest orchestrators — median turn ~1.7s — but with a long tail (P95 ~3.2s), driven mostly by LLM variance and telephony. Treat the sub-500ms figure as a best-case for the ElevenLabs-controlled stages, not the end-to-end phone experience.

[Heuristic] Practical latency levers, highest-impact first:

  • Pick a fast LLM. The single biggest, most variable slice is LLM first-token. A Flash-tier or Haiku-tier model can save 400–700ms over a frontier model.
  • Use eleven_flash_v2_5 for the voice unless you have a specific quality reason not to.
  • Keep the system prompt and per-turn context tight — long prompts inflate first-token time.
  • Co-locate: choose the region closest to your callers; telephony hops dominate global calls.
  • Mask unavoidable waits (tool calls, DB lookups) with a spoken filler ("Let me pull that up…") so dead air never exceeds ~1s.

Agent configuration

System prompt for a spoken conversation

[Heuristic] A voice-agent prompt is not a chatbot prompt. Spoken output is heard once, in sequence, with no scrollback, so:

  • Force brevity. Instruct one idea per turn, ~1–3 sentences. Long paragraphs are unlistenable and inflate TTS latency and cost.
  • Ban screen-only formatting. No markdown, bullet lists, tables, emojis, or URLs read aloud — they get spoken literally ("asterisk asterisk"). Ask for spelled-out numbers where clarity matters ("one-eight-hundred").
  • Give a persona and a scope. Role, tone, what it can and cannot do, and an explicit "if you don't know / it's out of scope, say so and offer transfer" clause.
  • Script the edges. First message, how to handle interruptions, silence, off-topic asks, and hostile callers.
  • Name the tools' triggers in prose. State when to call each tool, because the LLM decides tool use from the prompt plus tool descriptions.

[Documented] The dashboard splits this into a First message (the opening line, supports dynamic variables) and a System prompt. A recent platform field, trust_context, marks an agent as low (serving untrusted external callers — restrict tool access) or high (serving the owner — full tool access appropriate) (changelog, 2026, verified 2026-07-10).

LLM selection

[Documented, volatile — verified 2026-07-10] ElevenAgents supports a large, frequently-updated roster including:

  • OpenAI: GPT-5.5, GPT-5.4 (+ Mini/Nano), GPT-5.2, GPT-5.1, GPT-5 (+ Mini/Nano), GPT-4.1 (+ Mini/Nano), GPT-4o, GPT-4o Mini.
  • Google: Gemini 3.5 Flash, 3.1 Pro/Flash-Lite (preview), 3 Pro/Flash (preview), 2.5 Flash, 2.5 Flash-Lite.
  • Anthropic: Claude Opus 4.7, Sonnet 4.6 / 4.5 / 4, Haiku 4.5.
  • ElevenLabs-hosted open weights: Qwen3.6-35B-A3B, Qwen3.5-397B-A17B.
  • Custom LLM: bring your own OpenAI-compatible endpoint with stored credentials.

The exact list changes almost monthly; always read the live Models page rather than trusting a memorized roster.

[Heuristic] Choose by the job:

  • Latency-first, simple flows (FAQ, routing, quick info): a Flash / Flash-Lite / Haiku / Mini-tier model. Lowest first-token time.
  • Tool-heavy or reasoning-heavy flows (multi-step, disambiguation, function calling that must not misfire): a higher-intelligence model — ElevenLabs' own tool-use guidance recommends the stronger Gemini Flash, GPT-5-class, or Claude Sonnet 4.x tiers and specifically warns against the weakest Flash models for reliable tool calls.
  • Data-residency-constrained (EU): some older Gemini/Claude variants are unavailable under EU residency; OpenAI and custom endpoints remain available. Verify per-model.

[Documented] LLM cascading / fallback: ElevenAgents can fall back through a backup sequence when the primary model errors or times out, protecting the conversation from a single provider outage. A documented default cascade (verified via docs search 2026-07-10) runs Gemini 2.5 Flash → Gemini 2.0 Flash → Claude 3.7 Sonnet → Claude 3.5 Sonnet v2. [Heuristic] For production, keep cascading on and make sure every model in the chain is fast enough that a fallback doesn't wreck the call.

Voice and TTS model

[Documented, verified 2026-07-10] TTS model tradeoffs for agents:

Model ID Latency Languages Best for
eleven_flash_v2_5 ~75 ms ~32 Default for real-time agents. Best speed/quality balance, ~50% lower cost.
eleven_flash_v2 ~75 ms English only English-only, cost-sensitive.
eleven_multilingual_v2 Higher 29 Highest lifelike quality/emotion; long-form; when quality > latency.
eleven_v3 Higher 70+ Most expressive/dramatic; character work; generally not the low-latency real-time pick.
eleven_turbo_v2_5 Mid ~32 Deprecated for agents; superseded by Flash v2.5.

[Heuristic] Default to eleven_flash_v2_5. Escalate to eleven_multilingual_v2 only when a brand demands maximum warmth/expressiveness and can absorb the added latency. Higher-quality voice + model + LLM combined can visibly slow the turn — the dashboard itself warns of this.

[Documented] Voice settings that matter for agents: stability (lower = more emotional range and variation, higher = flatter/consistent), similarity boost (adherence to the source voice), and speaker boost (similarity at a slight latency cost). TTS output can be mp3, pcm, or ulaw — telephony typically needs ulaw 8kHz; SIP media runs G711 8kHz or G722 16kHz.

[Heuristic] Pick a voice whose native accent/language matches the caller base, keep stability moderate (very low stability can produce jarring swings across turns), and test the voice on your actual script — sample text flatters voices that stumble on domain jargon, phone numbers, and proper nouns.

Language and multilingual

[Documented] Agents can run in a fixed language or detect/switch languages; TTS covers 70+ languages overall, with Flash v2.5 at ~32 and Multilingual v2 at 29. [Heuristic] For a multilingual agent, confirm all three stages support the target language: ASR recognition quality, the LLM's fluency, and the chosen TTS model's coverage. A voice cloned from one language may carry an accent when speaking another — audition it.

Turn-taking and interruption

[Documented] ElevenAgents 2.0 uses a hybrid turn-taking model: a VAD plus deep-learning signals (filler words, prosody, rhythm, micro-pauses) to judge end-of-turn and handle barge-in. Configurable controls include turn eagerness (Eager / Normal / Patient) and a turn timeout from 1–30s (verified 2026-07-10).

[Heuristic]

  • Eager cuts response delay but risks interrupting slow or thoughtful speakers; good for brisk transactional flows.
  • Patient avoids cutting people off; good for support, elderly callers, or noisy lines, at the cost of a longer perceived pause.
  • Background noise (HVAC, cross-talk, call-center ambiance) can trip the VAD and cut the agent off mid-sentence — [Community/independent, disclosed] ElevenLabs publishes no false-interruption precision/recall metrics, so validate barge-in behavior on real telephony audio, not a quiet office mic.
  • A background_sound (ambient audio) option exists and an agent_response_complete client event now fires once when a turn is fully delivered — useful for reliable end-of-turn UI, wake-word gating, or message batching (changelog 2026).

Knowledge base and RAG

[Documented, verified 2026-07-10] Attach documents to an agent as grounded knowledge:

  • Sources: file upload (PDF, TXT, DOCX, HTML, EPUB; max 21MB/file), URL import (single page — it does not crawl linked pages), or pasted text.
  • Account limit: non-enterprise caps at 20MB / 300k characters total; enterprise expands via sales.
  • Config: documents live under conversation_config.agent.prompt.knowledge_base, each with type, name, id, and an optional usage_mode. RAG can be enabled so the agent retrieves relevant chunks rather than stuffing everything into context.
  • ElevenLabs does not publish the embedding model, chunk size, or retrieval internals.

[Heuristic] Split large manuals into focused documents, keep them current, and mine transcripts for questions the KB failed to answer. Put stable reference material in the KB; put per-user facts (account balance, order status) in dynamic variables or a server tool — never bake a customer's data into the KB.

Tools and function calling

[Documented] Three tool types:

  • Server / webhook tools — the agent makes an HTTP call to your backend for real-time data (DB lookups, order status) or authenticated actions (booking, refund). All secrets stay server-side.
  • Client tools — invoke JavaScript in the caller's app/browser: open a modal, navigate, update UI, read local state. Configure with type: "client", a name, description, parameters, and expects_response / "Wait for response" — when enabled, the agent pauses and appends the returned data to the conversation before continuing.
  • System tools — built-in actions: end call, transfer to another agent, transfer to a human/number, skip turn, detect language. (System tools cannot update dynamic variables.)

[Heuristic] Tool reliability depends heavily on the LLM. Give each tool a crisp description and each parameter a clear description and value_type so the model extracts arguments correctly. For flows where a mis-fired tool is costly (payments, cancellations), use a higher-intelligence LLM and add tool-call tests (below). Design tools to be idempotent where possible and mask their latency with a spoken filler.

Personalization: dynamic variables vs. overrides

[Documented, verified 2026-07-10]

  • Dynamic variables inject runtime values ({{user_name}}, {{account_tier}}) into the prompt, first message, and tools — the recommended way to personalize per conversation. Passed via conversation_initiation_client_data.dynamic_variables. System variables like system__caller_id, system__called_number, system__call_duration_secs, system__agent_id are auto-available but cannot be set/overridden by the client payload.
  • Overrides wholesale-replace the system prompt, first message, or other config at conversation start. Still supported but discouraged versus dynamic variables for maintainability.
  • Twilio personalization: on an inbound call ElevenLabs webhooks your endpoint with call info; you return conversation-initiation data (dynamic variables + overrides) to tailor that specific call.

Multi-agent workflows and transfers

[Documented] Agent Workflows are a visual editor for branching conversation graphs that route to specialized subagents, each with its own prompt, tools, and scoped knowledge base, and can hand off to a human. Two handoff mechanisms:

  • Agent-to-agent transfer (system tool): the parent hands the live conversation to a child agent when a described condition is met ("user asks about billing"). Full transcript is preserved; the transfer_to_agent call is stripped from the child's visible history so it continues seamlessly. Supports an optional delay and an optional spoken transfer message (silent if blank).
  • Transfer to human / number: routes to a person via the phone system, including warm-transfer into a conference with the caller.

[Heuristic] Prefer several small, single-purpose subagents over one giant prompt: each stays short (lower latency, fewer instructions to violate) and easier to test. Reserve human transfer for genuine escalation and always give the agent an explicit fallback path so a stuck caller is never trapped.

Channels and SDKs

[Documented, verified 2026-07-10]

  • SDKs: React (@elevenlabs/react), JavaScript/Web (@elevenlabs/client), Swift (iOS), Kotlin (Android), React Native, plus Python/TypeScript server SDKs. A CLI (@elevenlabs/cli) manages agents as code.
  • Transport: voice conversations default to WebRTC; text-only defaults to WebSocket. The React startSession takes signedUrl, conversationToken, or agentId and infers transport from mode.
  • Widget: an embeddable HTML snippet keyed by agent ID for zero-code website deployment.
  • Telephony: native Twilio integration (inbound + outbound, no infra change) and SIP trunking compatible with most providers (Twilio, Vonage, Telnyx, Plivo, Bandwidth, Sinch, Infobip, RingCentral, Exotel, and other standards-compliant trunks). SIP supports digest or ACL/IP auth, TLS signaling, and SRTP media; codecs G711 8kHz or G722 16kHz. Batch outbound calling is available for campaigns.

Authentication (do not leak your key)

[Documented] Never put the ElevenLabs API key in client code. For authenticated agents:

  • WebSocket → your server calls the REST API to mint a signed URL (expires ~15 min); the client connects with it.
  • WebRTC → your server mints a conversation token; the client connects with it.

[Heuristic] Always run authenticated agents in production so anyone can't rack up minutes on your agent ID, and gate the token/signed-URL endpoint behind your own auth. Public/unauthenticated agents are fine only for open demos.

Testing, simulation, and evaluation

[Documented, verified 2026-07-10] ElevenAgents has a built-in test framework, runnable from the dashboard Tests tab, the CLI (elevenlabs agents test <agent_id>), or the SDK (run_tests / runTests, with repeat_count for probabilistic runs):

  • Scenario / LLM-eval tests — a simulated user drives a conversation; an LLM scores the agent against success criteria you write (include tone expectations plus success and failure examples).
  • Tool-call tests — assert the agent calls the right tool with the right args, validated by exact match, regex (e.g. ^ORD-[0-9]{8}$), or LLM semantic check.
  • Simulation tests — full or partial multi-turn simulations (a partial one starts mid-conversation to unit-test a sub-flow), with tool mocking and custom evaluation, scored against an outcome statement you define.
  • Probabilistic testing — run a test N times; badges show green (100%), amber (≥80%), red (<80%) pass rates, grouping failures by reason. Because LLM output is stochastic, single-run "it worked once" is not a pass.
  • Create test from conversation — turn a real failed call into a regression test in one click.

[Documented] Post-conversation evaluation criteria (success rubrics) and data collection (structured extraction) run automatically on live conversations using the same criteria as simulation, feeding analytics.

[Heuristic] Build a suite covering the happy path, each tool, each transfer condition, and the hostile/off-topic/silent edges; run it before every prompt or model change. Prompt and model swaps routinely regress behavior that "obviously" worked.

Analytics

[Documented] The platform provides conversation search, transcripts, success-evaluation results, extracted data-collection fields, A/B testing, and performance metrics for monitoring at scale. [Heuristic] Watch median and tail latency, interruption/barge-in rates, tool error rates, transfer-to-human rate, and success-criteria pass rate — the tail and the transfer rate reveal production pain the averages hide.

Pricing, concurrency, and burst

[Documented, volatile — verified 2026-07-10] ElevenAgents bills by call minutes, separate from the shared TTS credit pool; the LLM and telephony are billed on top by usage. Representative self-serve tiers (included minutes / concurrent calls):

Plan Included minutes Concurrent calls
Free 15 4
Starter 75 6
Creator 275 10
Pro 1,238 20
Scale 3,738 30
Business 12,375 40

Overage ≈ $0.08/min standard, $0.16/min burst; text messages ≈ $0.003 each. Burst pricing, when enabled, lets you exceed your concurrency cap up to ~3× at the doubled rate. There is no cap on how many agents you can create; the binding constraint is concurrent calls.

[Heuristic] Concurrency, not minutes, is what breaks at scale — a spike of simultaneous inbound calls hits the ceiling instantly and extra callers get rejected or burst-billed. For real call-center volume (hundreds of simultaneous calls) you need an enterprise concurrency arrangement; do not assume a self-serve plan absorbs a marketing blast. Model both levers: total minutes for the monthly bill, concurrent calls for the busy-hour peak.

Safety, consent, and disclosure — a synthetic voice talking to real people

This is a legal and ethical surface, not a nicety. [Documented / regulatory]

  • AI disclosure: Disclose that the caller is talking to an AI. Under the EU AI Act (Article 52 transparency for synthetic/deepfake audio), and analogous emerging rules, artificially generated voice interacting with a person must be disclosed. [Heuristic] Put the disclosure in the first message ("Hi, I'm an AI assistant from …"), not buried later.
  • TCPA (US): ElevenLabs' synthetic voices count as an "artificial or prerecorded voice." Marketing calls need Prior Express Written Consent; informational calls need prior express consent. Call only 8am–9pm in the recipient's local time zone; scrub against the National DNC Registry and your internal DNC list; honor opt-outs promptly (within 10 business days) and offer an automated opt-out. Consent cannot be a condition of purchase. Many US states impose stricter rules — get legal counsel.
  • GDPR / biometrics: Voice is generally biometric data under GDPR. Processing it typically needs an Article 9 lawful basis (usually explicit consent) plus a DPIA (Article 35). For EU/biometric processing use the enterprise tier and a signed DPA.
  • Voice cloning consent: If the agent uses a cloned voice, you need documented, explicit written consent from the voice owner covering AI/synthetic voice-model creation and generated speech — cloning carries its own compliance posture beyond stock TTS.
  • Prohibited uses: impersonation, fraud, evading voice verification, unauthorized robocalling, harassment, voter suppression, and similar. Don't build them.

[Heuristic] Bake compliance into config: disclosure in the first message, an easy "talk to a human" transfer, a recording/consent notice where required, timezone-aware scheduling for outbound, and DNC checks before dialing. Treat "is the user OK talking to a bot, and did they consent to the call?" as a design requirement, not an afterthought.

Common failure modes

[Heuristic, from documented behavior]

  • Reads formatting aloud — markdown/emojis/URLs in output. Fix: forbid them in the prompt; request plain speech and spelled-out numbers.
  • Rambling turns — no brevity constraint. Fix: cap turns at 1–3 sentences in the prompt.
  • Gets cut off in noise — VAD false-triggers. Fix: Patient turn-taking, test on real telephony audio, tune the timeout.
  • Sluggish replies — frontier LLM + high-quality voice. Fix: Flash-tier LLM + eleven_flash_v2_5, tighten the prompt, add spoken fillers for tool waits.
  • Tools misfire or don't fire — weak LLM or vague tool descriptions. Fix: stronger LLM for tool-heavy flows, sharpen descriptions/params, add tool-call tests.
  • Leaked API key / unauthenticated agent — key in client code. Fix: signed URLs / conversation tokens minted server-side; authenticated agents in production.
  • Concurrency wall at launch — self-serve concurrency too low for peak. Fix: size the plan to busy-hour simultaneous calls, enable burst, or go enterprise.
  • No AI disclosure / no consent — regulatory exposure. Fix: disclose in the first message; secure TCPA/GDPR consent before deploying.

Complete example (labeled example, not a required formula)

Intent: an authenticated, low-latency support agent for a mid-size e-commerce brand, deployed as a website voice widget with a phone fallback, that can look up order status and escalate to a human.

  • Trust context: low (external callers).
  • LLM: a fast Flash/Sonnet-tier model with high enough tool reliability for order lookups; LLM cascading left on.
  • TTS: eleven_flash_v2_5, moderate stability, brand voice matched to the customer base's accent.
  • First message (also the AI disclosure): "Hi, you're speaking with an AI assistant from Northwind Support. I can check an order or connect you to a person — how can I help?"
  • System prompt (essentials): role = order-status support; keep replies to 1–2 spoken sentences; never read URLs, markdown, or emojis; spell out order numbers and tracking codes; if the order can't be found or the request is out of scope (refunds, complaints), apologize briefly and transfer to a human; call get_order_status whenever the user references an order.
  • Tools: server tool get_order_status(order_id) → your backend; system tool transfer_to_number for escalation. order_id parameter has a clear description and a regex-checkable format.
  • Personalization: {{customer_first_name}} and {{account_tier}} injected as dynamic variables at session start; system__caller_id used to pre-lookup the account on phone calls.
  • Knowledge base: shipping-policy and returns-policy documents (as focused files, RAG on) for general questions — not per-customer data.
  • Turn-taking: Normal eagerness, ~7s timeout; test barge-in on recorded phone audio.
  • Channel: React SDK widget over WebRTC using a server-minted conversation token; Twilio number for inbound phone with a personalization webhook.
  • Testing: scenario test (empathetic handling of a "where's my order" with success + failure examples), tool-call test asserting get_order_status fires with a regex-valid order_id, and a partial simulation for the escalation branch — each run with repeat_count ≥ 5.
  • Expected result: sub-second-feel replies for cached info, a spoken filler while get_order_status runs, clean handoff to a human on out-of-scope asks.
  • Likely failure modes: order-lookup tool misfires on ambiguous input (mitigate with a stronger LLM + clear param description); noisy callers cut off the agent (mitigate with turn-taking tuning); launch-day concurrency spike (size the plan / enable burst).
  • Variations: swap the widget for pure telephony via SIP trunk; add a billing subagent and route via Agent Workflows; run outbound reminders as a batch campaign — which then triggers full TCPA consent + calling-window obligations.

Primary sources (verified 2026-07-10)