ElevenAgents Architect: The ElevenLabs AI That Repairs Voice Agents

ElevenAgents Architect is a new AI assistant from ElevenLabs that reads your live voice agents’ conversations, finds what went wrong, drafts a fix, tests it, and proposes it for approval. The one-minute launch video pitches it as agents that “improve themselves.” The documentation describes something more careful and more useful, and this post covers both.

TLDW

You ask Architect about an agent that is already taking calls. It investigates, recommends a change, and builds it. With your approval the change can be piloted on a share of live traffic before a full rollout. It also raises problems you have not asked about, and it can build a new agent from a spoken description and some documents. As of October 6, 2026, the documentation says Architect is in alpha and free to use through October.

Thoughts

“Your agents can now improve themselves” is the headline, and it is not quite what ships. According to the documentation, every change Architect makes is staged as an unpublished draft and is never applied to live callers directly. A person publishes it, or Architect opens a merge proposal for a teammate to approve. So the agents do not improve themselves. They propose their own improvements, with the evidence attached, and a human decides. That is the right design for something answering a company’s phone, and it is a better product than the slogan suggests. An agent that silently rewrote its own instructions after a bad call would be much harder to trust than one that shows a diff.

The part that matters is validation. Architect writes tests and simulations for its own change and runs them before asking you to ship, and the video says its unprompted suggestions arrive “already pressure tested against simulations.” This is the same idea that has made coding agents useful over the past year: an agent that can check its own work can iterate without a person in the middle. The open question is who checks the checker. Tests written by the system that wrote the fix will tend to confirm the fix. The real test is the pilot on live traffic, and for a voice agent that means real customers hear the experiment. Keep the pilot share small and decide in advance what result would make you roll it back.

Two details in the documentation deserve more attention than they will get from a launch video. Architect conversations are not covered by ElevenLabs’ Zero Retention Mode, and the docs advise against putting personal or sensitive information in messages to it. Any team in healthcare, finance or another regulated field that turned that mode on for a reason should read that line before pointing Architect at call transcripts. And the price is unknown. The alpha is free and does not count against agent minutes or LLM cost, which is a good reason to try it this month and a poor basis for planning around it.

Key Takeaways

Architect is the assistant built into ElevenAgents, the ElevenLabs platform for voice and chat agents. You describe what you want in plain language and it edits the agent for you.
Its main job is improving an agent that is already deployed. The documented loop is investigate, change, validate, propose.
It works on the same objects you can edit by hand: the system prompt, procedures, workflow, tools, knowledge base, guardrails and tests. Its edits show up in the same draft, diff and version history as yours.
Nothing reaches live callers until a person publishes the draft. Architect can open a merge proposal for a teammate and suggest a small traffic split to try the change on live calls.
It is proactive. The video says it “catches issues you’ve missed,” and the docs describe handoffs from Spotlight, which produces a weekly summary, alerts and suggested investigations.
It can build a new agent from a description. The video’s example is a single spoken request for an appointment-setting agent, with documents attached and a note that half the callers prefer Spanish.
Limits listed in the docs: it is not available in every data residency environment, attaching files requires a Creator plan or higher, idle sessions end after 5 minutes (15 on Enterprise), and conversation analysis is capped at 100 conversations per user per minute.
Architect chats are private to the person who started them, and they are not covered by Zero Retention Mode.
Alpha status: free through October 2026, with no charge against agent minutes or LLM cost. Pricing after that has not been announced in the docs.

Detailed Summary

What the Video Shows

The video runs 62 seconds and makes four claims in order. You can ask Architect anything about your live agents, and it will find the root cause of a problem, recommend a fix and build it. Once you approve, it can pilot the change on live traffic before rolling it out broadly. It proposes changes without being asked, already tested against simulations. And it builds new agents from context you give it, “the way an expert would,” to standards you set.

How It Works According to the Docs

Architect reads an agent’s configuration, real and test conversations, test results, insights, alerts and triage tickets. It edits the configuration as a staged draft, writes tests and simulations for the change, and runs them in the same conversation. There are no hidden settings: it uses the same configuration that the dashboard exposes. It is reachable across the ElevenLabs app for product questions and navigation, and from inside a specific agent, where it can read and change that agent.

Approval and Rollout

The documentation is firm that changes are drafts until a human publishes them. For teams, Architect can open a merge proposal so a colleague reviews the change, in the way a pull request works for code. For risk, it can suggest a small traffic split so the new version handles only part of the live calls. The docs do not spell out percentage controls or rollback steps on that page.

Availability, Privacy and Price

Architect is in alpha through October 2026 and free during that period. It is not offered in all data residency environments. Because its conversations fall outside Zero Retention Mode, the docs say to avoid sharing personal or sensitive information with it. For pricing after the alpha, the docs point to the pricing page.

Notable Quotes

“Your agents can now improve themselves.”
ElevenLabs launch video, the opening line and the claim to read most carefully

“It finds the root cause, recommends a fix, and builds it for you.”
ElevenLabs launch video, on what happens when you ask about a live agent

“Architect doesn’t wait to be asked. It catches issues you’ve missed and proposes changes.”
ElevenLabs launch video, on proactive suggestions

The video is short enough to watch twice. Watch the launch video here, then read the documentation before turning it on for a production agent.