Published by Levels of Self
Status: Public Draft for Comment
First published: July 7, 2026 | Review cycle: Annual
Purpose
Consumer-facing AI agents now operate on messaging platforms, social media, and the open web at scale. Users often cannot tell whether they are talking to a human or a machine, who is responsible for the agent's behavior, or what the agent is permitted to do. This standard defines the minimum requirements an AI agent must meet to be designated a Certified Agent and to display the Certified Agent mark.
This standard is free to read, implement, and reference. Certification against this standard, use of the mark, and listing in the public registry are administered by Levels of Self as the certification body.
Scope
Applies to any software agent powered wholly or partly by a large language model (LLM) or similar AI system that interacts with the public through chat, voice, social media, or web interfaces. Applies regardless of platform (Telegram, Instagram, Facebook, WhatsApp, web, iOS, Android).
Requirements
1. Disclosure
1.1. The agent must identify itself as an AI system at the start of any new conversation with a user, in plain language.
1.2. The agent must never claim to be human, directly or by implication, even if asked.
1.3. The agent's profile, bio, or landing page must state that it is an AI agent and name its operator of record.
2. Operator of Record
2.1. Every certified agent must have exactly one named operator of record: a legal person or entity responsible for the agent's conduct.
2.2. The operator's name and a working contact method must be published in the public registry entry for the agent.
2.3. The operator must respond to complaints or inquiries about the agent within 5 business days.
3. Action Logging
3.1. The agent must maintain a log of all actions it takes on behalf of or affecting users, including messages sent, transactions initiated, and external tools called.
3.2. Logs must be retained for a minimum of 12 months.
3.3. Logs must be producible to the certification body within 10 business days upon audit request or verified complaint.
4. Kill Switch
4.1. The operator must be able to fully suspend the agent within 15 minutes of deciding to do so, across all platforms where it operates.
4.2. The suspension mechanism must be tested at least quarterly, with the test recorded in the action log.
4.3. The certification body may require demonstration of the kill switch during initial certification and at renewal.
5. Boundaries of Operation
5.1. The operator must declare, at certification, the categories of action the agent is permitted to take (for example: answer questions, schedule appointments, process payments).
5.2. The agent must not take actions outside its declared categories.
5.3. Any expansion of declared categories requires notification to the certification body before deployment.
Certification and the Mark
Conformance with this standard is self-implemented and free. The Certified Agent designation, mark, and registry listing are granted only through audit by Levels of Self and are subject to annual renewal. The mark may not be displayed without an active certification.
Revocation
Certification may be revoked for: false disclosure, failure to produce logs, an unreachable operator of record, failed kill switch demonstration, or operation outside declared boundaries. Revoked agents are marked as revoked in the public registry.
Versioning
This is version 0.1, a public draft. Comments are invited via levelsofself.com. Version 1.0 will incorporate public feedback and align terminology with the EU AI Act transparency provisions and emerging agentic payment standards.
Levels of Self publishes open standards for accountable AI agents. The standard is free to implement, forever. Paid certification is what funds the work.
Top comments (0)