ElevenLabs

ElevenLabs explains multimodal input in ElevenAgents: images and PDFs arrive as native model messages on web, mobile, and WhatsApp. Covers file-backed versus inline representation, setup, and carrying context across sessions.

Lessons from Admiral on taking voice agents to production in insurance: a bounded first use case, compliance involved from day zero, deterministic logic for provable steps, human escalation, and simulation tests with 1 to 100 percent rollouts.

ElevenLabs now hosts open-source LLMs for its agents platform. GLM 4.5 Air targets reasoning and tool calling at lower cost, while Qwen3-30b-a3b targets low latency, running beside its speech and turn-taking models.

ElevenLabs details a DTMF system tool in ElevenAgents that lets voice agents call out and navigate third-party IVR phone trees. It covers the timing, depth, and provider challenges and the outbound workflows this unlocks.

A recap of an ElevenLabs workshop on making voice agents sound human. Covers four traits (rhythm, mirroring, context, identity), a restaurant reservation demo driven by prompt design, and seven levers including emotional tags, eagerness, pronunciation dictionaries and custom voices.

Guidance on feeding enterprise knowledge to voice agents: choosing direct context inclusion or RAG by knowledge base size, preparing documents for retrieval, and automating bulk uploads and refreshes.

ElevenLabs CLI v1 exposes the full API as subcommands with structured JSON output and a dry-run mode. It adds agents-as-code: pull agents to local config, preview diffs, push to production, and use branches to separate dev and prod. Built for coding agents.

A latency-budget walkthrough for cascaded voice agents: how time-to-first-audio accumulates across capture, endpointing, STT, LLM, TTS, network and playback, plus streaming, codec and regional measurement tactics.

ElevenLabs added voicemail detection as a system tool for its conversational agents. The LLM spots automated greetings and the agent then leaves a custom message or ends the call, reducing wasted usage on outbound campaigns.

ElevenLabs launches Spotlight, an observation layer for ElevenAgents. It auto-groups conversations by topic, scores them against plain-language evals, tracks sentiment, detects metric anomalies, and suggests the next improvement for deployed voice and chat agents.

An architecture guide to sub-200 ms streaming speech-to-text covering transport choice, interim versus committed transcripts, end-pointing, in-stream features, audio formats and benchmarking, with code for the capture and forwarding path.

A hands-on guide to layering voice onto existing agent frameworks (LangGraph, Google ADK, CrewAI, LlamaIndex) through a Custom LLM proxy that streams OpenAI-compatible SSE chunks, with FastAPI examples and per-framework handling.

ElevenLabs engineers describe the orchestration engine behind ElevenAgents: multiple models per answer pipeline with under 100ms overhead, how conversation context is built, and the tradeoff between independent agents and multi-agent workflows.

Explains how Scribe v2 Realtime streaming works: partial versus committed transcripts, choosing a commit strategy, and using server-issued single-use tokens for client-side connections in a React live translation demo.

ElevenLabs makes Scribe v2 Realtime available in its agents platform. The streaming speech-to-text model transcribes in 30-80ms and reports the lowest word error rate among low-latency ASR on FLEURS across 30 languages, based on internal benchmarks.

ElevenLabs made Scribe v2 Medical generally available, a speech-to-text model fine-tuned for clinical dictation. Its published results on Corti and Eka benchmarks show lower overall and medical-term error rates than base Scribe v2, OpenAI and Deepgram Nova-3 Medical.

ElevenLabs launched ElevenAgents Architect, a built-in expert that builds agents from provided context, finds root causes in live agents, proposes fixes pressure-tested in simulation, and can pilot changes on live traffic before broad rollout. Based on the post and demo transcript only.

ElevenLabs launches Architect inside ElevenAgents: a conversational expert that reads transcripts and Spotlight findings, proposes prompt, workflow and tool changes, generates simulations, and saves versioned drafts.

ElevenLabs outlines its layered safety design for agents: prompt-injection guardrails on input, workflow-gated sensitive actions, parallel output validators, pre-launch simulations, and post-launch evaluation. Vendor-specific, but a concrete pattern.

ElevenLabs adds Procedures to its voice agent platform (alpha). They come in structured and free-form modes, load only when triggered, and can be drafted from imported SOP documents.

ElevenLabs ships synthetic voice detection inside ElevenAgents for enterprise customers, with broader rollout later this month, and joins the Meta and Sierra-led Personal Agent Protocol group defining how personal agents identify themselves to businesses.

ElevenLabs brings Eleven v4 Turbo into its ElevenAgents platform, pairing the roughly 100ms TTS model with its transcription and turn-taking models, supporting 90+ languages and mid-call language switching.

ElevenLabs Agents gained a system tool that generates keypad tones across telephony stacks, so voice agents can work through IVR phone trees on outbound calls. It opens fully automated phone workflows such as benefits verification and rebooking.

ElevenLabs explains how it cut RAG latency by racing query-rewrite models, including self-hosted Qwen variants, and falling back to the raw message after a one-second timeout. Median latency fell from 326ms to 155ms, with p95 down to 426ms.