ElevenAgents Architect: Free Through October 2026, and the Proactive Part Isn’t Live Yet
ElevenLabs launched ElevenAgents Architect on 6 October 2026 — an AI assistant built into ElevenAgents that builds and fixes your voice and chat agents while you talk to.
ElevenLabs launched ElevenAgents Architect on 6 October 2026 — an AI assistant built into ElevenAgents that builds and fixes your voice and chat agents while you talk to it in plain language.
The 67-second launch video makes a specific promise near the middle: that Architect “can also be proactive, reasoning across thousands of conversations to spot opportunities you’d otherwise miss,” with each one coming back “as a proposal, with the fix already built and validated in simulation.”
ElevenLabs’ own documentation says that part is not live yet. Under Known limitations:
No scheduled or background work yet. ElevenAgents Architect only works while you’re in a conversation with it. It doesn’t yet scan agents on a schedule, send reports, or open proposals on its own. Scheduled automations are coming soon.
That is not a gotcha so much as a roadmap feature being shown in present tense. But if the proactive part is why you were interested, it is the first thing to know. Here is what Architect actually does today, what the alpha costs, and the three other things worth checking before you point it at a production agent.
What Architect actually does
Architect is an assistant inside ElevenAgents that works on the same objects you would edit by hand — the system prompt, procedures, workflows, tools, knowledge base, guardrails, tests, branches and merge proposals. The docs are explicit that there is no parallel system: “It has no separate hidden configuration. Every change it makes appears in the same draft, diff, and version history as a change you made yourself.”
It is built around a four-step loop:
- Investigate. Read real conversations, Spotlight insights, alerts, triage tickets and failing tests to find what is going wrong and why.
- Change. Edit the prompt, a procedure, a tool, the knowledge base or guardrails. Configuration changes are staged as an unpublished draft and, in the docs’ words, “never applied to live callers directly.”
- Validate. Write tests and simulations for the change and run them inside the same conversation, before you are asked to ship anything.
- Propose. Hand the change to you. You publish the draft, and Architect can open a merge proposal and suggest a small traffic split to try it on live calls.
It also builds new agents from a description, answers product questions and walks you through the dashboard. You can use it by voice or chat in the product, and from Claude, Claude Code, ChatGPT, Cursor and Grok Bot.
It is free, but only until the end of this month
Architect is in alpha, and the docs put a date on it: it is “free to use during the alpha period, which runs through October 2026.” During the alpha, Architect conversations do not count against your workspace’s agent minutes or LLM cost.
After that, the docs say only: “Check the pricing page for terms after the alpha ends.” No post-alpha price has been published. So if you want to evaluate Architect on someone else’s meter, you have the rest of October.
Three limits apply during the alpha:
- Attaching files in the composer requires a Creator plan or higher.
- An idle session ends after 5 minutes, or 15 minutes on Enterprise plans. You can continue from chat history.
- Conversation analysis is capped at 100 analyzed conversations per user per minute.
Rollout is also partial. The docs note that some surfaces, including the per-agent Architect tab and the triage queue, “are being rolled out gradually and may not appear in every workspace yet.” If you open ElevenAgents and cannot find it, that is expected.
The privacy line worth reading twice
This one is easy to miss, and it matters most to exactly the people ElevenLabs is selling agents to — healthcare, financial services, government. From the docs:
Architect conversations are not covered by Zero Retention Mode (ZRM). Avoid sharing personal or sensitive information in messages to Architect.
Zero Retention Mode is the Enterprise control that stops ElevenLabs retaining conversation data. Your agents’ calls can be covered by it; your chats with Architect about those calls are not. Since investigating an escalation naturally involves pasting in what a customer said, that is a real operational gap, not a theoretical one.
Architect chats are also private to you. Teammates cannot open them, even when the chat was started from a shared triage ticket. If you want a colleague to see the outcome, you share the merge proposal or a triage ticket comment, not the conversation.
“Nothing goes live without your approval” — with an asterisk
The launch material says nothing ships without you. For agent configuration, that holds: changes are staged in your draft and only you can publish a draft.
But there are two things worth knowing. First, Architect has approval modes. The default is Approval required, where it asks before actions that affect live traffic or shared resources — merging a branch, changing a traffic split, updating a tool. There is also an Auto-approve mode, and the docs are clear that in that mode it does not ask before those actions.
Second, merging is deployment. Straight from the FAQ: “Does merging a proposal deploy it? Yes. There’s no separate publish step after a merge. The merged version serves live traffic immediately.”
There is a neat safety side effect in how permissions work. Architect runs as you, with your permissions — it cannot do anything your role does not allow. Because it acts as you, you are the author of any proposal it opens, and authors cannot review their own proposals. So a teammate with editor access has to approve it. Workspace admins can merge without an approval.
Architect is not the ElevenLabs MCP server
Worth separating, because both let an AI assistant change your agents and they behave differently. The docs: “No, but they call the same APIs. Architect runs in the dashboard, stages changes in drafts, and uses its own approval modes. The MCP server lets external assistants such as Claude and Cursor manage your agents, applies changes directly, and relies on your client’s tool approvals.”
Applies changes directly is the operative phrase. If you have been managing ElevenAgents through the MCP server in Claude or Cursor, you have not been getting the draft-and-review safety net that Architect gives you in the dashboard.
Spotlight does the watching, Architect does the fixing
ElevenLabs now has two agent-improvement products and they are easy to confuse.
| Product | Role | What it does |
|---|---|---|
| Spotlight | Observe and recommend | Monitors conversations; weekly summary, suggested investigations, real-time alerts, configuration recommendations |
| Architect | Build and fix | Investigates a question or a Spotlight finding, changes the agent, writes and runs tests, prepares the change for review |
Spotlight cards hand off to Architect with buttons like Analyze with Architect and Investigate with Architect. Until scheduled automations arrive, Spotlight is the thing that notices, and Architect is the thing you then open.
What else it cannot do yet
- It cannot review merge proposals. It can open one and suggest reviewers, but approving, commenting, merging and closing are human-only.
- Some resources are read-only to it: MCP servers, secrets, integrations, channels, alerts and phone number assignments.
- Dashboards built in the chat are not saved — the data lives only in the browser tab where it was built.
Who should try it this month
If you already run a deployed ElevenAgents agent with real conversation volume, this is worth the rest of October. The investigate-change-validate loop is aimed squarely at an agent that is already live and underperforming, and the fact that it writes the tests for its own change is the genuinely useful part. The docs even suggest a good habit: ask Architect to show the new tests failing without the change and passing with it.
If you are building your first agent, Architect will build one from a description, but you are evaluating a tool whose price after this month is unpublished.
And if you work somewhere that needs Zero Retention Mode, read that privacy line before your team starts pasting transcripts into it.
The verdict
Architect is a real product doing a real job, and the draft-plus-simulation design is more careful than most “AI edits your config” features. The gap between the launch video and the documentation is about tense, not truth — the proactive agent is coming, it just is not here yet.
Take the free month. Point it at an agent that is already live. And decide what you think before the pricing page fills in.
You can try ElevenAgents Architect here, or read the ElevenLabs announcement in full.
All quotes, limits and dates in this post were read off the ElevenLabs blog, the ElevenAgents Architect documentation and the launch video on 7 October 2026. Architect is in alpha and these terms are explicitly stated to change after October 2026.