Agent Governance Card
State governments are moving past chatbots into agentic AI: software that acts. Agents can use tools, touch data, receive and send files, and complete multi-step tasks with autonomy. Yet the IT governance machinery states rely on today, including intake, security review, procurement, and audit, wasn’t built for software that acts. The introduction of agents in government presents governance challenges.
A first central step to address them is to ensure visibility and transparency on what agents do in public organisations. What is an agent allowed to do? What did it touch? Who answers for it? How does a person step in? And what does it cost to leave? These are questions that cannot and should not be avoided.
The Agent Governance Card was developed by The Agentic State and the Center for Civic Futures to help teams address these questions. It exists to assess and record information, bringing structured, comparable transparency to agentic deployments that interact directly with the public or materially influence a decision affecting the public or state resources.
Who uses it. Most often a product team, the people who build or configure the agent. A CIO’s office, a procurement officer, or an auditor may also fill one in. When a vendor is engaged, the accountable public agency fills it.
When. Most often before an agent gets production access to state data, tools, or users. Also earlier, as a checklist during development, or later, for systems already live. Once filled, the card is revisited on a rolling basis.
What it delivers. A snapshot of one deployment: capabilities, data used, who answers for it, how a person steps in, and more. From the answers, the card computes an Impact Level that routes the deployment to proportionate obligations.
The backbone of the card is its schema. It builds on the governance frameworks that matter most for public deployments, including OMB’s memoranda on AI use and procurement, the UK’s Algorithmic Transparency Recording Standard and NIST’s AI Risk Management Framework, and on frontier research on agents: identity and delegation, autonomy and reversibility, the cost of leaving a vendor.
The schema is composed of twelve dimensions, spanning what an agent does and touches, who answers for it, its risks, and more. Across them, 47 fields: for each, the plain-language question a team answers, the values an answer can take, and the rationale behind the field. Fourteen fields, plus the Impact Level, form the public layer, the set of information that would ideally be made public.
The schema can be filled in two ways. Teams can work through it by hand, or fill it interactively by running the skill, which acts as the interviewer: it asks the questions one at a time, helps with the hard ones, computes the Impact Level, and emits the filled card. The skill installs in any assistant that supports the open Agent Skills standard, and can be opened directly in the repository of the agent being built, where it reads the answers from the specs and code already there.
You can explore the card, the schema and worked examples on the Agent Governance Card website, get the skill and templates from the GitHub repository, or download the full document (PDF, 1.5 MB), which brings everything together with the rationale behind every field, guidance on what evidence to use and who should fill in the card, and answers to common questions.
The Agent Governance Card is released under a CC BY 4.0 licence. Share any feedback via hello@agenticstate.org.