ElevenAgents Architect: Alpha Access, Features, Limits and Pricing

Quick answer: ElevenAgents Architect is a new built-in expert inside ElevenLabs’ agent platform that can investigate agent performance, diagnose likely root causes, change configurations, write and run tests, and prepare a versioned change for a human to review. Architect is currently in alpha, is rolling out gradually, and ElevenLabs says Architect conversations are free through October 2026 without consuming workspace agent minutes or LLM cost.

The important part is that Architect is not another end-user voice agent. It is an agent for the people who build and operate agents. Instead of manually moving from analytics to prompts to simulations to settings, a team can ask a question in natural language and let Architect investigate across those layers.

What ElevenAgents Architect is designed to do

ElevenLabs positions Architect as an expert partner for building and improving agents. It understands the configuration surfaces inside ElevenAgents, including knowledge bases, prompts, workflows, voice settings, procedures, guardrails, tools and simulations.

A typical request might start with a broad operational question such as “Why are refund conversations failing?” Architect can inspect the relevant data, identify a likely cause, modify the agent configuration, create tests and run them before presenting a proposed change for review.

Architect’s core workflow

The current product flow can be summarized in five stages:

1. Investigate: analyze a question or a finding from Spotlight against the underlying data.
2. Diagnose: connect the observed behavior to prompts, tools, procedures, knowledge or other configuration.
3. Change: edit the agent configuration to address the issue.
4. Test: write and run tests or simulations against the proposed change.
5. Review: prepare a versioned change so a person can inspect it before applying it.

This review step is important. Architect can do the investigative and experimental work, but its output is prepared for human review rather than silently replacing an agent configuration without visibility.

How Architect connects with Spotlight

ElevenLabs’ Spotlight surfaces findings about agent behavior. Architect can receive those findings directly through actions such as Analyze with Architect or Investigate with Architect.

That creates a useful feedback loop: monitoring identifies a pattern, Architect investigates it, a fix is tested, and the resulting change can be reviewed as a version. Teams no longer have to translate every analytical finding into a separate manual debugging session.

What can Architect change?

According to ElevenLabs, Architect understands the available configuration options across the platform and how they interact. That includes:

• system and task prompts;
• knowledge-base content and retrieval setup;
• workflows and procedures;
• voice and conversation settings;
• guardrails and behavioral rules;
• tools and integrations;
• simulations and tests.

The value is not simply that it can edit a prompt. Many agent failures happen because several settings interact. A good refund flow, for example, can fail because the knowledge base is stale, a tool returns incomplete data, the escalation rule is too aggressive, or a test case does not cover a common user phrasing.

ElevenAgents Architect pricing during alpha

Architect is currently free during the alpha period through October 2026. ElevenLabs says Architect conversations do not count against a workspace’s agent minutes or LLM cost during that period.

That does not mean every ElevenLabs service is free. The no-cost treatment applies to Architect conversations during the alpha. Normal agent traffic, telephony, models and other billable platform features continue to follow their own pricing rules.

Current availability and rollout

Architect is not necessarily visible in every workspace yet. ElevenLabs says surfaces such as the per-agent Architect tab and triage queue are being rolled out gradually.

There are also environment and plan-specific constraints:

• Architect is not yet available in every isolated data-residency environment.
• Attaching files in the composer requires a Creator plan or higher.
• An idle Architect session ends after five minutes on standard environments.
• Enterprise idle sessions can remain open for up to 15 minutes.
• Conversation analysis is subject to platform limits, including a documented rate limit on analyzed conversations.

Example: fixing a low refund-resolution rate

Suppose an agent handles customer refunds, but the resolution rate has dropped. A manual investigation could require checking transcripts, failure tags, the refund policy, tool responses, prompts and simulated calls separately.

With Architect, the operator can begin with the business symptom. Architect can trace the symptom to likely causes, inspect the relevant configuration and data, propose a change and test it. The operator then reviews the version rather than manually reconstructing the entire chain.

This is where the feature differs from ordinary “AI assistant for settings” experiences. The promise is not merely a recommendation; it is an investigation-to-test workflow grounded in the actual agent configuration.

Where Architect can be most useful

• Post-launch optimization: diagnose why successful test behavior degrades in real conversations.
• Prompt regressions: identify changes that caused an agent to become less consistent.
• Tool failures: determine whether errors come from the model, instructions or an external action.
• Knowledge gaps: find questions that the current knowledge base cannot answer reliably.
• Guardrail tuning: reduce unnecessary refusals without weakening important policies.
• Simulation coverage: create tests for failure patterns found in production data.

What teams should verify before accepting a change

Even when Architect’s diagnosis is convincing, treat the prepared version like any other production configuration change. Check that the proposed fix solves the target problem, does not introduce regressions in unrelated flows, and respects privacy, legal and business rules.

For high-impact agents, maintain a stable evaluation set so every Architect-generated change is tested against the same critical scenarios. That makes the tool most useful as an accelerator inside an existing quality process rather than a substitute for one.

Architect vs a normal AI assistant

Why this release matters

Voice-agent platforms are moving from “build a bot” toward continuous operations: observe conversations, identify failure patterns, fix configuration and prove that the fix works. Architect compresses several of those steps into one conversational workflow.

The timing also makes the alpha notable for teams already experimenting with voice AI. AVARIXO recently covered Amazon Nova 2.5 Sonic, another example of rapid development in conversational AI infrastructure.

FAQ

Is ElevenAgents Architect free?

During the alpha period through October 2026, ElevenLabs says Architect conversations are free and do not consume agent minutes or LLM cost.

Is Architect available to everyone?

No. The rollout is gradual, and some surfaces may not yet appear in every workspace. Some isolated data-residency environments are also not supported yet.

Can Architect edit an agent?

Yes. It can make configuration changes as part of an investigation and prepare a versioned change for human review.

Does it run tests?

Yes. Testing and simulation are central to the workflow: Architect can write and run tests before presenting a proposed change.

Does Architect replace human review?

No. The current workflow explicitly prepares changes for review. Teams should still validate important production changes.

Sources

Primary sources: ElevenLabs — Introducing ElevenAgents Architect and ElevenLabs Architect documentation. Checked October 7, 2026.