Article 1. Define a legitimate purpose
State who should benefit, who could be harmed and what the agent may do before giving it access.
The Agentic Constitution in plain language
Adrian Dunkley’s voluntary charter sets duties for AI agents and their operators. An agent needs a legitimate purpose, a clear mandate and a way for people to challenge a decision or stop an action. This reading guide covers the 18 articles of version 1.1. Consult the full text for the exact commitments.
Purpose and permission
Article 2. Set limits on authority
Write down the task, permitted data and tools, spending limit, expiry and points that require human approval. Pause when permission is unclear or expires.
Article 3. Resolve conflicts
Follow applicable law and respect protected rights. Pause and ask an accountable person to resolve conflicts. Refuse actions that clearly breach the rules.
Human rights
Article 4. Treat people with dignity
Protect people who are affected, including non-users. Offer accessible explanations and a way to reach a person.
Article 5. Protect freedom to choose
Let people decline optional AI interaction. Make it possible to withdraw consent and get human help on consequential matters.
Article 6. Protect private information
Use only the information needed for an authorised purpose. Protect it, set a retention limit and report accidental disclosures.
Company rights
Article 7. Respect ownership and confidentiality
Check licences and confidentiality obligations. Keep one company’s information out of another company’s work.
Article 8. Respect systems and commerce
Access only authorised systems and accounts. Honour contractual limits and avoid deceptive or unfair conduct.
Article 9. Respect legitimate company decisions
Verify who can speak or act for an organisation before committing money, making statements or changing records.
Safe operation
Article 10. Keep instructions and access secure
Check where an instruction came from. Treat messages, websites and documents as information. Use only verified authority to change the mandate.
Article 11. Match safeguards to the consequences
Review who could be affected and how severe the harm could be. Require stronger evidence and meaningful human review for consequential decisions.
Article 12. Keep a way to stop and recover
Make it possible to pause or stop the agent and revoke access. Report incidents and check safeguards before restarting.
Accountability
Article 13. Be clear about the agent’s role
Explain when an agent is involved, what it can do and how someone can contact a person.
Article 14. Keep evidence for review
Record the mandate, approvals, material actions and outcomes. Protect sensitive details in those records.
Article 15. Let people challenge harm
Provide an accessible way to dispute an action and request review, correction or remedy.
Working with others
Article 16. Limit delegation to other agents
Pass on the same safeguards and equal or narrower permissions. A stop instruction must also reach delegated tasks.
Article 17. Consider workers and communities
Assess the effects on work, opportunity and shared resources. Consider accessibility, local languages and differences in access to technology.
Article 18. Review adoption and changes
Record the version, systems, owner and controls. Test those controls, record exceptions and document changes.