# ElevenAgents Architect for live voice agents > ElevenAgents Architect drafts and tests fixes for a voice agent in production, so your engineers review changes instead of hunting for them. Published: 2026-10-19 Author: Raja Vemuri Tags: AI/ML, Applied AI Canonical: https://www.techfabric.com/blog/elevenagents-architect-improve-voice-agents --- If your company runs a voice agent in production, keeping it working takes engineers you'd rather have on the roadmap. Every week brings calls it handles badly, and each fix starts with someone reading transcripts, editing the prompt and testing again by hand. ElevenLabs has added [Architect](https://elevenlabs.io/docs/eleven-agents/operate/architect) to ElevenAgents, its platform for voice agents, to take over most of that cycle. It reads a deployed agent's conversations and test results, drafts a change to its prompt, procedures, tools, knowledge base or guardrails, and runs tests and simulated conversations against that change. Changes to the agent's configuration wait as an unpublished draft until a person on your team publishes them. Your engineers review fixes with the evidence attached, and stop spending their week hunting for them. ## What it changes for the team that owns the agent For a team running a live voice agent, the slow part of improving it moves. Finding the problem and writing the first fix becomes the assistant's job, and judging the proposal becomes the team's. That is a real saving when the backlog of odd calls is longer than anyone's week. It also puts a new decision in front of the team. Somebody has to own the review of these proposals and say what evidence is enough to publish one. The risk sits in how changes reach callers. In the dashboard, configuration changes are staged as a draft, and an auto-approve mode that skips approval for merging branches and changing traffic splits is off by default. From an external assistant such as Claude or Cursor, connected through ElevenLabs' hosted MCP server, updates are applied directly, and updating an agent publishes a new version on the branch you name, or on `main` if you name none. Architect is free during an alpha that runs through October 2026, and its conversations don't count against agent minutes or LLM cost in that period. ElevenLabs points to its pricing page for the terms after that, so a plan that assumes it stays free is a guess. ## A lender's collections line Take a regional lender whose voice agent handles inbound calls on its collections line. Spotlight, the monitoring view in ElevenAgents, raises a suggested investigation. Callers who ask to move a payment date get passed to a person far more often than the team expects, and the human queue backs up on the days after statements go out. The team could open Architect from that Spotlight suggestion, which passes the investigation and its context in. Architect would read the conversations behind it, look for the cause of the escalations, and could find, say, a procedure that never covers date changes. It would then draft a revised procedure as an unpublished draft and write a simulation test, with a borrower asking to move a payment as the scenario and a turn limit to end the run. The success condition would say the agent offers only the options the lender's policy allows and hands off anything outside them. The team would set that simulation to mock the servicing tools, so the run never touches the real loan system. The collections lead and a compliance reviewer would read the diff and the simulation result before anyone publishes. Architect could then suggest a small traffic split, so the new procedure meets real callers on part of the line first, and Spotlight keeps watching the same calls afterwards. The compliance reviewer has one more constraint to hold the team to. Architect conversations aren't covered by Zero Retention Mode, so borrower names, balances and account details never go into a message to it, even while investigating a single call. ```mermaid %% caption: A Spotlight suggestion becomes a draft procedure, a simulation with mocked servicing tools, a human review and a small traffic split, and Spotlight keeps watching. flowchart TD A["Spotlight suggests an investigation"] --> B["Architect reads the conversations"] B --> C["Draft procedure change, unpublished"] C --> D["Simulation test, servicing tools mocked"] D --> E["Collections lead and compliance review the diff"] E --> F["Publish to a small traffic split"] F --> A class E accent ``` ## Starting on one deployed agent with tests We'd start with one live agent that already has tests, and one real signal from it, such as a failing test or a Spotlight suggestion. The docs say Architect is most useful for improving an agent you've already deployed, and that is where we'd judge it, because real calls give a proposal something to be measured against. We'd write the success condition of the deciding simulation ourselves. A passing simulation is weak evidence until a person has decided what passing means. ## Writing the team's standards into agents.md Next we'd write the team's standards into the agent's `agents.md`, the document Architect receives at the start of every conversation about that agent. The [customization docs](https://elevenlabs.io/docs/eleven-agents/operate/architect/customization) give the kind of rule we mean, such as every procedure change needing a simulation test, or never changing the voice without asking. There is no workspace-wide rules document yet, so each agent carries its own copy. ## Keeping the external route off main We'd keep the dashboard's draft and approval model exactly as it ships, and treat the [external assistant route](https://elevenlabs.io/docs/eleven-agents/operate/architect/external-assistants) with more care. Before anyone changes a production agent from their editor, we'd set the update and merge tools to ask first in the client and have the assistant work on a branch, so nothing lands on `main` by accident. ## What sits behind the agent's tools Architect can't fix what sits behind the agent's tools. If the loan servicing data is wrong, a better procedure reads the wrong balance more politely. That part is integration work, and our [agents work](/ai/agents) covers the pattern we prefer there, with narrow tools, recorded actions and a person approving anything with consequences. ## The first investigation to run Pick the Spotlight suggestion that costs your human queue the most, and write its success condition yourself before Architect drafts anything. Then judge it on whether the change you publish still holds after a week of real calls on that line. ## Frequently asked questions ### What is ElevenAgents Architect? ElevenAgents Architect is an AI assistant built into ElevenAgents that investigates an agent's conversations and test results, drafts changes, and validates them with tests and simulations. ElevenLabs describes its output as proposals for you to review, and you stay responsible for approving every change before it reaches callers. ### Does Architect change a live voice agent without approval? Not by default in the dashboard, where configuration changes are staged as an unpublished draft and a person publishes them. Auto-approve mode can skip approval for actions such as merging branches and changing traffic splits, but it is off by default. Changes made from an external assistant through the hosted MCP server are applied directly, so that route needs its own safeguards. ### Can a team use Architect from Claude or Cursor? Yes, Claude, Claude Code, ChatGPT and Cursor connect through ElevenLabs' hosted MCP server and sign in with OAuth. Coding assistants can also install the ElevenLabs plugin, which bundles the server with Architect's own skills. Agent context and personal preferences aren't sent automatically on that route, so the assistant knows only what you tell it and what it reads with its tools. ### Is ElevenAgents Architect free? It is free during the alpha, which runs through October 2026, and its conversations don't count against agent minutes or LLM cost in that period. ElevenLabs points to its pricing page for the terms after the alpha ends, so check that page before you plan beyond October.