Short answer: Govern an AI agent system by giving a named person, not another machine, authority over what the agents may do and what they release. A practical setup has five parts:
- One controlled access point. Route every agent’s access to AI platforms through a single gateway that holds the platform keys.
- A signed scope before work starts. A named governor approves each task’s scope, risk class, and platforms, and the system runs only what was signed.
- Human decisions at consequential checkpoints. The governor sees the full answers and accepts, modifies, or rejects them. For outputs too fast to review one by one, the governor signs the policy that authorizes that class of outputs.
- Signatures no agent can make. Approvals use a passkey on the governor’s own device, the system holds only public keys, and high-risk work adds a second governor who countersigns.
- A tamper-evident record with stated limits. Chain the record by hash, and remember that a signature shows what was signed, not what was read.
What does governance add to an agent system?
Governance adds a named person who can stop the agent and answers for what it does. Without that person, an agent system can be well engineered and still be automation checking automation. Basil C. Puglisi, MPA, frames the test in his analysis of why agentic AI was always going to fail: “If no, the system is Responsible AI (automation without a named human bound to the output) regardless of how autonomous it is. If yes, the system is AI Governance regardless of how autonomous it is.”
The Checkpoint-Based Governance page explains why the human cannot be replaced by more machines: “No platform count, confidence score, or level of agreement substitutes for human arbitration.” Basil’s book Governing AI: When Capability Exceeds Control goes further in Appendix B, “Checkpoint-Based Governance (CBG v5.0),” which treats identical agreement across AI platforms with no dissent as a signal to look harder rather than as proof of accuracy.
An organization can choose how much autonomy to allow. The page defining what an AI agent is describes three agent models, from one where a human signs a charter and the machine acts on every step after it, to one where the human performs every step. Choosing among them “is itself a checkpoint decision, recorded with its risk rationale.”
Where does the named human sit?
The named human sits at the points where an agent’s work would become a decision. Basil’s paper from the definition problem to running code shows one way to build that into software, in a program the author calls the CBG Gateway. Its design offers a pattern other agent systems can adapt.
- One door to the AI platforms. The gateway “is the only program that holds the AI platform keys, so agents and automations cannot reach the platforms except through it.”
- A signed scope before work starts. “An agent proposes a task. A named governor establishes it,” with authority to rewrite its scope, raise its risk class, choose the platforms, and sign. The system “then runs only what the signed manifest names.”
- Full answers at each checkpoint. Each prompt goes to three to seven AI platforms, “every answer is kept in full as it arrived, with no filtering and no merging,” and the governor “accepts, modifies, or rejects it.”
- No advance without a decision. “No role advances, and no result is released, without a verified signature from the governor named in the signed manifest.”
The paper calls this the hybrid model, where “The machine carries the dispatch, the collection, the routing, and the record at machine speed,” while “a named human decides with the complete answers in view and signs for the decision.”
Speed does not remove the human; it moves the checkpoint. For outputs too fast to review one by one, the Named-Human Test in why agentic AI was always going to fail places the person upstream, where “the named human signs the policy that authorizes the class of outputs.”
What does a governor sign, and what does the signature establish?
A governor signs every act of authority with a credential bound to the person’s own device. In the gateway, “Every act of authority requires a passkey signature made on the governor’s own device after a fingerprint, face check, or device PIN, and the gateway holds only public keys, so no AI and no agent can approve anything.” Passkeys are built on FIDO standards, and the FIDO Alliance explains that they “replace passwords with cryptographic key pairs for phishing-resistant sign-in security” and “are designed so that there are no shared secrets.”
High-risk work adds people. “A high-risk task names a second governor, who must countersign the final text,” and “Two people stand behind every new passkey after the first, so no single operator can enroll a governor alone.” The record is chained by hash, and “Every release writes a receipt that anyone holding the gateway’s public key can check offline.”
The signature has limits, and the paper states them. “The gateway establishes what was signed, not what was read.” It “governs only the AI use that passes through it, and someone can still paste a prompt into a chat window outside it.” In The Minds That Bend the Machine, Chapter 16 reviews Stuart Russell and presses on this gap, asking how anyone can tell a real decision from a signature on something the signer could not fully assess.
Status matters too. The code for both coded agent models is Alpha: “Neither is deployed, and in the build record neither has yet called a live AI provider,” and “Neither has had a professional security review.” The gateway has “run only against canned platforms.” The pattern is offered as one possible path to adapt, not a product to install.
For an organization with agents already running, Puglisi Consulting’s AI Deployment and Agent Review “compares what was promised with what was delivered and finds where the deployment broke from its own plan.” It is one option, and the checks above can start inside the organization.
Questions Readers Ask
What is the first step in governing AI agents?
Make one controlled point the only way agents reach AI platforms, and control their access to company systems and data the same way, so every action passes where it can be governed. Then inventory each agent, define what it may do, and assign a named person with authority to approve its scope and stop it when needed.
Should every agent action require human approval?
Not necessarily. Low-risk, routine actions can run within limits a named person has signed. Consequential actions need a checkpoint where a person decides with the full evidence in view. For actions too fast to review individually, a person signs the policy that authorizes that class of actions.
Why use passkeys for agent governance signatures?
Passkeys bind a signature to the person’s own device and a biometric or PIN check, and they use key pairs with no shared secrets. If the system holds only public keys, no AI or agent can produce a valid approval, which keeps the authority with the named person.
What can a governor’s signature not prove?
A signature shows that an enrolled governor’s passkey approved the item after the device verified the user, but it cannot show that the person read or understood the material. It also covers only work that passed through the governed system, so agent use outside that system needs separate controls.
When should a second governor countersign?
High-risk work benefits from a second named person who must countersign the final result. Requiring two people to enroll any new signing credential also prevents a single operator from creating a governor alone. Both steps spread authority so no one person can approve consequential output unchecked.
Related Questions
- Who is accountable when an AI agent makes a mistake?
- Compare the most effective AI governance frameworks for managing agent-based systems
Sources
FIDO Alliance. (n.d.). FIDO passkeys: Passwordless authentication. https://fidoalliance.org/passkeys/
Puglisi, B. C. (2025). Governing AI: When capability exceeds control (2nd ed.). https://basilpuglisi.com/governing-ai-when-capability-exceeds-control/
Puglisi, B. C. (2026, June 4). Why agentic AI was always going to fail. basilpuglisi.com. https://basilpuglisi.com/why-agentic-ai-was-always-going-to-fail/
Puglisi, B. C. (2026, October 13). HAIA agents: From the definition problem to running code. basilpuglisi.com. https://basilpuglisi.com/haia-agents-running-code/
Puglisi, B. C. (2026). The minds that bend the machine. https://basilpuglisi.com/minds-that-bend-the-machine/
Puglisi, B. C. (2026, October 13). What is an agent? Code, human, or both. basilpuglisi.com. https://basilpuglisi.com/agents/
Puglisi, B. C. (2026, September 11). What is Checkpoint-Based Governance? Human authority. basilpuglisi.com. https://basilpuglisi.com/cbg/
Puglisi Consulting. (n.d.). Advisory reviews. https://puglisiconsulting.com/advisory-reviews/
#AIgenerated using the HAIA Ecosystem | CC BY-NC-SA 4.0
Free for personal, educational, and noncommercial research use with attribution. Commercial exploitation, paid productization, and enterprise commercialization require separate permission and licensing.
Suggested citation: Puglisi, B. C. (2026). I need advice on setting up governance for AI agent systems. basilpuglisi.com.
Leave a Reply