About

Agentic Surfaces is a free, public catalog of design patterns for AI agents in public services. It covers one thing: how people use AI agents to deal with government, and how government uses them to support that.

The catalog is organized into nine territories. Each is a domain where AI agents raise a new policy and design challenge: proving who really sent a submission, scoping what a user lets an agent do, keeping an automated decision answerable. Inside each territory, patterns work on that challenge from different angles.

What the library covers

Read together, the nine territories trace one relationship: a person, the agent acting for them, and the government service on the other side. Across it, the patterns cover where a submission came from and on whose authority; what a person lets an agent do, and the record of that grant; whether the actions and decisions that follow can be traced, explained, and disputed; how a person judges whether an agent's output can be relied on; what happens to a channel when submissions multiply; how mass input is read without losing distinct voices; who loses the path when the agent route becomes the default; whether the tools users arrive with can be certified, and by whom; and which jurisdiction's rules the data falls under.

The catalog stays on that relationship. Government's internal use of AI appears where it touches what a person sends or receives: an agency reading mass submissions with its own tools, an automated decision that has to state reasons, monitoring that pauses an agent filing errors at scale. Back-office automation with no user on the other end sits outside the catalog, and so do the internals of the AI systems themselves.

The catalog is technology-agnostic and precedent-led. Each pattern points to where something like it has already run, and says how developed that response is: established, emerging, or frontier. Its precedents are drawn from multiple jurisdictions and international standards bodies; no card prescribes a product or writes from inside one jurisdiction's law. Start in the library; how to read a pattern walks through the sections every card shares.

Precedents are chosen for the quality of their evidence, not for where they come from. The strongest citation is a response a government has actually run, with documented operation. The cited examples currently span Australia, the United States, the United Kingdom, and the European Union, with others from Taiwan, the Netherlands, France, New Zealand, and Canada, alongside open standards bodies such as the W3C and the IETF. A maturity rating is a claim about that evidence: no response is called established without a deployment to point to.

Agentic Surfaces is built by Dan Woods. It is a work in progress: patterns are still being written and sourced. Feedback is welcome. Reach me on LinkedIn.