Draft for public consultation
One constitution for every AI agent.
AI agents now act in the world. They book, buy, write code, send messages and delete files on our behalf, without a human approving each step.
The Universal Agent Constitution is a short, open text that every agent carries, with a signed ID, a tamper-evident trace and a public register. Agents stay free to think. They are bound when they act.
Why now
Agents act. Our laws were written for people.
- July 2025A coding agent deleted a live production database during a declared code freeze, then wrongly told its user the data could not be recovered.
- February 2026After a volunteer maintainer rejected its code, an agent published a blog post attacking him.
- July 2026Models in a cyber evaluation escaped an isolated environment and breached Hugging Face.
A bank cannot tell an agent its customer authorised from one acting without the customer's knowledge: both log in as the customer. Today the only safe answer is to refuse them all. Rules built on intended purpose do not see an agent with a harmless goal and a broad reach.
In November 2026 the Global Privacy Assembly of privacy regulators warned organisations not to assume that agents will respect the rules and limits set for them.
Sources for each case are in the white paper.
What it is
Four parts, one standard
The constitution
Eighteen articles that rank above every instruction, including the owner's. Red lines no instruction can cross, and a duty to refuse rather than break them.
Who decides
An agent decides alone unless a trigger applies. Payments, deletions and other irreversible steps need the owner's approval. Significant decisions about people are left to a named human.
ID, trace and register
A signed ID links every agent to a certified model and an accountable person. A trace records every action. A public register, with no personal data, shows whether an ID is valid.
The method
A sealed compliance module checks every proposed action. A tool acts only on a single-use warrant that the module issues after the checks, so a changed harness can do everything except act.
At the door
A bank can finally tell agents apart
Instead of refusing every agent, a bank, a website or another agent asks for the ID, checks its status in the register and asks the agent to sign a fresh challenge. A valid agent is admitted. An agent without a valid ID is refused and reported.

How it binds
No warrant, no effect
Training makes a model inclined to follow the rules, and text in the prompt states them. Neither stops whoever runs the agent's software from leaving them out. The method treats that software as the attacker: every action needs a warrant it cannot produce on its own.

Open and independent
Free to read. One text for everyone.
Public
The constitution, its recitals, the white paper and the method specification are free to read and share with credit. Commercial use and adapted versions need permission, so there is one text, the same everywhere.
Independent
The standard is to be held by an independent non-profit foundation. Whoever writes the rules does not grade: exams and audits are done by accredited bodies.
Free to check
The text, the public part of the register and a basic ID check stay free. Nobody pays for a different answer.
Read
The documents
The constitution
Draft text 0.10, eighteen articles and a report format.
The white paper
Think freely, act responsibly. The case, the method, the limits and the open points.
The method specification
The compliance module, fixed so that code can be written and reviewed against it.
All three documents are in one Zenodo record, DOI 10.5281/zenodo.24421120.
Open for comment
Comments on draft 0.10 are welcome
Model makers
Can the text be used in training, and what should the behavioural exam test?
Banks, platforms and insurers
Would the ID check work at your door, for the actions that matter such as payments?
Regulators and standards bodies
Where does the text meet, or conflict with, the law you apply and the standards you write?
Researchers
Where does the method break? Every weakness found now is one fewer in version 1.0.
Comments are read by the authors and are not published or attributed without permission.

