ElevenLabs releases Architect alpha to repair AI agents by chat
In some approval modes of an external AI tool, a change to an agent can go live with no check.
Claimed, not confirmed
ElevenLabs, an AI audio company, told on the 19th that it released an alpha (test) of ElevenAgents Architect. Users make and change voice and chat agents when they tell the tool in voice or text. No code and no menus are necessary. Changes in the dashboard are kept as drafts that a manager must examine. The company gave no numbers or test results.
How it works
Architect knows how knowledge, prompts, workflows, voice and guardrails (limits on the steps of an agent) work together. A manager can name a problem, for example a low refund rate. The tool reads the conversations and test results, finds the cause and shows a change. It can examine thousands of conversation records at one time and tests each change in a simulation.
The proof
ElevenLabs gave no test results. Carl Hughes of Deacon Group tells that Architect gives speed and control to voice agent teams in regulated fields.
Open points
Architect also works from external tools:
A change from such a tool can go live with no check. It is not known how companies control this.
Insight: Platforms open their controls to rival AI tools
On October 14, OutSystems made Agent Experience work with Claude Code, Cursor, Codex and Kiro. ElevenLabs does the same with Claude, ChatGPT and Cursor. This reads as platform vendors that put their controls in the tools of the users, to keep the customer. The weak point is approval, which is in the external tool.
Sources
Posted
