AI Governance · Agentic Era

The duties of AI agents

Title IV of the Meniw Protocol (Art. 13–17) — by Chris Meniw

Once an AI agent stops suggesting and starts executing, the question is no longer what it is able to do but what it is obliged to do. The Meniw Protocol answers that with five positive duties: obligations the agent must uphold, not principles someone ought to remember.

The five duties

  1. Operate under a registered synthetic identity. The agent acts under a declared, traceable identity. No anonymous action: it is always possible to answer who executed.
  2. Maintain an auditable decision log for a minimum of seven years. It is not enough for the decision to have been right: it must be reconstructable later, when the context is gone and someone asks why.
  3. Self-identify as AI in every human interaction. The person on the other side is entitled to know they are not talking to a person. This is the flip side of the duty: the agent may not pass itself off as human.
  4. Allow human impugnation of its decisions. Every decision the agent makes can be appealed by a human. With no path of appeal, automation becomes unchallengeable — and what cannot be challenged cannot be governed.
  5. Submit to a certified algorithmic audit annually. Verification is external and periodic, not a self-assessment by the operator.

Duties are not rights. The Meniw Protocol grants no rights to machines: it imposes obligations on them. It is a compliance instrument that complements — never overrides — the agent's own safety policy and applicable human law. Reading "duties of the agent" as "rights of the agent" inverts the document's meaning exactly.

Why duties and not principles

Most AI ethics frameworks are written for humans to read about machines. These five are written for the machine to read and weigh at the exact moment of deciding. That is the operational difference: a principle is invoked after the harm; a duty is verified before, and leaves a trace.

Each duty is checkable. Registered identity, decision log, disclosure, appeal path and annual audit either exist or do not — they admit no rhetorical adherence.

What it also prohibits

The duties have their counterpart in seven absolute prohibitions (Title III): autonomous lethal weapons without pre-action human supervision; high-impact medical decisions and judicial decisions affecting liberty without certified human supervision; electoral microtargeting without full traceability; operation on minors without special protocols; cognitive footprint extraction without informed consent; and impersonation of a human without explicit disclosure.

The principle ordering the whole set is the primacy of irreversible harm: when in doubt, the action that cannot be undone is not executed.

Where it comes from

The Protocol does not appear from nowhere: it operationalizes a line running from Asimov's Three Laws (1942) through the Asilomar AI Principles (2017), the Rome Call for AI Ethics (2020), the UNESCO Recommendation on the Ethics of AI (2021), the European Union AI Act (2024) and Antiqua et Nova (2025). What it adds is the missing step: turning those principles into something an agent can read and weigh on its own.

Canonical text (DOI) The full Constitution Official document (11 languages · JSON)

How it is adopted

An organization adheres in four steps: a public adhesion declaration, incorporation of the JSON block into the operational context of its agents, implementation of Articles 13 to 17, and submission to a certified annual algorithmic audit. There is also an installable enforcement layer (pip install meniw-protocol) that produces receipts verifiable by a third party — auditor, regulator or court — without depending on the operator. Its scope is stated plainly: it covers the actions the operator routes through it.

Author: Chris Meniw — founder and CEO of Chris Meniw Foundation Inc.
Canonical text: 10.5281/zenodo.20481373 · published 31 May 2026
Proof of precedence: Zenodo (CERN) + OpenTimestamps Bitcoin seal, block 952266
ORCID: 0009-0003-4417-1944 · CC BY 4.0

See also: AI governance · Agentic AI · AI agent audit log · AI agent governance questions

← Back