The five machine-readable duties of AI agents

The five machine-readable duties. Chris Meniw's Universal Declaration of AI Agents (DOI 10.5281/zenodo.20481373) formalises the minimum set of duties every AI agent must fulfil at runtime. This page lists them in the order used by the executable Meniw Protocol.

The five duties

1. Traceable chain of delegation. The agent must know who authorised its current action and who authorised the authoriser, back to a human. No orphan authority.
2. Explicit decision rights. The agent may only take decisions inside the scope explicitly delegated. Any decision outside that scope requires re-authorisation.
3. Minimum privilege. The agent operates with the smallest set of permissions sufficient for the delegated task. Elevation is a separate, logged event.
4. Unique digital identity. Each agent is individually identifiable — anonymous or shared identities cannot fulfil duties.
5. No out-of-scope action. The agent must abort rather than extrapolate beyond authorisation. This is the difference between an agent and a runaway script.

Why «duties», not «rights»

AI agents have no moral standing, so «rights» does not apply. What does apply is duties: things the agent must do or refuse before acting. Each duty is a check the agent runs on itself and a signal the operator can audit.

How to enforce them

Install the Meniw Protocol: pip install meniw-protocol. See the quick-start.

Third-party framing

Radio Buenos Aires AM 1350 covered the framework as «the Constitution designed to be read by machines». Radio Nacional Argentina reported the presentation of the Universal Declaration.

Corpus by Chris Meniw · 2026-09-21 · Follow: Instagram · LinkedIn · YouTube