Exploring Governance-to-Runtime Control Mapping with ACS

[JoieC-GRC-Analyst]

Hello ACS maintainers and community,

I’d like to propose a small practitioner-focused use case exploring how AI governance requirements could be translated into runtime agent controls and observable evidence using the Agent Control Standard (ACS).

The proposed case study would use the New Brunswick Responsible AI Framework (NB-RAF), a practitioner-developed responsible-AI governance framework, as a source of example governance requirements.

The goal is not to propose NB-RAF as part of ACS or to suggest that ACS can establish organizational AI-governance compliance.

Instead, I’d like to explore the following question:

Can a governance requirement be translated into an enforceable and observable runtime control using ACS?

Proposed approach

If the community considers this useful, I would start with a small crosswalk of selected governance requirements:

Governance requirement → control objective → ACS mechanism → runtime decision → observable evidence → identified gap

I would then test the mapping against one concrete high-risk agent scenario, such as an organizational agent that can retrieve internal information and invoke external tools.

For example:

Governance requirement
Human approval is required before a high-impact external action.

→

Runtime control requirement
The agent’s tool invocation must be evaluated before execution.

→

ACS interaction
Evaluate the relevant agent/tool interaction and produce an appropriate control decision.

→

Evidence
The decision and resulting action should be observable through the ACS event/trace model.

The intent would be to test the mapping rather than assume that every governance requirement has an existing ACS representation.

Potential contribution

If the mapping produces useful results, the work could potentially become:

• a real-world use case;
• documentation;
• security analysis/feedback; or
• a governance-oriented example/profile.

If the analysis identifies a requirement that cannot be represented or enforced cleanly using the current ACS capabilities, I would document that as a gap rather than proposing a specification change prematurely.

Why I am asking first?

I saw that ACS asks contributors to seek feedback before investing significant effort and explicitly welcomes documentation, real-world use cases, and security analysis.

Before developing the crosswalk, I’d therefore appreciate feedback from the maintainers/community:

Would this governance-to-runtime mapping be useful to the ACS project, and if so, are there particular ACS mechanisms or use cases you would recommend using as the starting point?

Thanks for considering the idea. I’m happy to scope the work around the areas that would be most useful to the ACS community.