With Architect, ElevenLabs automates part of the AI agent lifecycle.

ElevenLabs launches ElevenAgents Architect, an assistant that can build, analyze, test and improve conversational agents through natural language.

ElevenLabs is adding another layer to its conversational agent platform. With ElevenAgents Architect, released in Alpha on October 6, 2026, the company is introducing a built-in assistant designed to build, analyze and improve agents through a written or spoken conversation.

The idea is to move part of the configuration process into natural language. A team can ask why certain refund requests are being escalated to human operators, look for ways to improve resolution rates, or request a new evaluation set. Architect can then inspect the agent and, when needed, retrieve its conversations, test results and other ElevenAgents data to investigate what is happening.

Its role goes beyond offering recommendations. Architect understands the different configuration layers inside ElevenAgents, including prompts, knowledge bases, workflows, voice settings, procedures, guardrails, tools and simulations. Once it identifies a potential issue, it can prepare a change, write related tests and run simulations before presenting the result to the team. ElevenLabs is effectively positioning Architect between conversation analysis and the actual modification of an agent.

The system can also reason across large numbers of conversations to identify recurring issues or opportunities that teams might otherwise miss. Instead of manually reviewing individual interactions, Architect can consolidate signals, propose an adjustment and evaluate it against simulated conversations. The decision to deploy that change ultimately remains with the user.

ElevenLabs has built several control mechanisms around that process. By default, actions affecting live traffic or shared resources can require approval. A Plan mode restricts Architect to research and preparing proposed changes, while an Auto-approve mode gives it significantly more autonomy. Its capabilities are also tied to the user's existing permissions, meaning someone with read-only access cannot use Architect to modify an agent they are not authorized to edit.

Changes can be staged in separate versions before they are published, allowing teams to inspect, test or revert them. Organizations can also provide agent-specific context through an `agents.md` file and personal working preferences through `user.md`. These documents can define an agent's purpose, tone, internal conventions and testing requirements that Architect should follow.

Architect can also start from scratch. Teams can describe the experience they want to create, attach supporting documents and ask it to build an agent around that context. ElevenLabs is therefore extending agent creation beyond technical specialists, while leaving more complex rollout strategy, integrations and architecture to dedicated teams.

The same approach is beginning to extend beyond the ElevenLabs interface. The company provides a hosted MCP server for working with ElevenAgents from environments including ChatGPT, Claude, Claude Code and Cursor. ElevenLabs also provides Architect-related skills for supported development assistants, using OAuth authentication and the same underlying agent management APIs.

ElevenAgents Architect is currently available in Alpha, with some parts of the experience still rolling out progressively across workspaces. ElevenLabs says Architect is free during the Alpha period through October 2026, with its conversations currently excluded from workspace agent minutes and model costs.

With Architect, ElevenLabs is moving beyond simplifying the initial creation of voice agents. The company is starting to automate part of their ongoing lifecycle: observing production conversations, investigating failures, preparing changes, testing them and submitting those changes for review. That shifts agent management closer to a continuous improvement process rather than a configuration that teams periodically revisit.