Agent Versioning and Configuration
ElevenLabs introduced agent versioning capabilities, allowing developers to fetch and update specific agent configurations using new version_id and branch_id query parameters. The Get agent endpoint now supports these parameters for version-specific retrieval, while the Update agent endpoint accepts branch_id for branch-scoped modifications. This enables better version control and branch-based development workflows for agent configurations.
RAG and Tool Expansion
A new search_documentation built-in system tool provides retrieval-augmented generation (RAG) capabilities with multi-source document retrieval. Developers can configure multiple sources via MultiSourceConfigJson, SourceConfigJson, and SourceRetrievalConfig schemas, with customizable result merging strategies. Additionally, MCP (Model Context Protocol) tool support is now available as a new tool type alongside existing webhook, client, and system tools, expanding agent extensibility options.
Content Moderation and Safety
Content guardrails have been added to agent configurations with configurable thresholds across seven categories: sexual content, violence, harassment, self-harm, profanity, religion/politics, and medical/legal content. The new ContentGuardrail and ContentConfig schemas allow fine-grained control over content moderation behavior. An expressive_mode boolean field enables agents to leverage the new v3 conversational model for more natural interactions.
Enhanced Conversation and Testing Features
A new endpoint (GET /v1/convai/users) lists users who have had conversations with agents, supporting pagination and filtering by agent, time range, and other criteria. Phone number transfers now support post_dial_digits for sending DTMF tones after connection. Agent testing was restructured into three distinct types: llm (response evaluation), tool (tool-call verification), and simulation (full conversation simulation) with dedicated schemas for each type.
SDK and Pronunciation Dictionary Updates
Python and JavaScript SDKs updated to v2.36.0 with support for all new features. The JavaScript SDK added overloaded convert signatures to speechToText.convert() for improved type inference. Widget packages (@elevenlabs/convai-widget-core and @elevenlabs/convai-widget-embed) now display agent tool usage status and long-running tool call badges. The Get pronunciation dictionary endpoint now returns full rule details in a rules array, and folder deletion support was added to the knowledge base API with recursive deletion capability via the force parameter.