SYNTROX.ai
Blog

NIST AI RMF for AI Agents: Govern, Map, Measure and Manage as Concrete Controls

5 min read

Read for agents, the NIST AI RMF's four functions become twelve concrete controls. GOVERN: an agent inventory, named owners, and plans for retirement and provider failure. MAP: a written mandate, oversight set per action, and every tool mapped. MEASURE: independent evaluation, production monitoring and tested safe failure. MANAGE: a release gate, a stop path with owners, and incident response.

On this page

The NIST AI Risk Management Framework, AI RMF 1.0 (NIST AI 100-1), was published in January 2023 and does not discuss AI agents. It does not need to. Its core is written as outcomes rather than techniques, and outcomes translate.

Three things are worth knowing before mapping it. The framework is voluntary, and NIST says the actions under its functions "do not constitute a checklist". NIST's AI RMF page now states that AI RMF 1.0 is being revised as part of the White House AI Action Plan, so treat what follows as a reading of 1.0. And the companion Generative AI Profile (NIST AI 600-1, July 2024) adds suggested actions for generative systems, and lists autonomous agents among the threats whose likelihood and magnitude should be assessed.

How to apply the NIST AI RMF to AI agents

The core has four functions, each broken into categories and subcategories. GOVERN is cross-cutting and informs the other three; NIST expects most users to start with MAP once governance is in place, then continue to MEASURE or MANAGE. Below are three agent controls under each function, with the subcategories they answer to. The subcategory numbers matter: they are what lets a risk committee trace a control back to the framework.

GOVERN: who owns each AI agent and its risks

1. An agent inventory. GOVERN 1.6 asks for mechanisms to inventory AI systems. The Generative AI Profile suggests that inventory entries record the underlying foundation models and their versions, and human oversight roles (GV-1.6-003). For an agent, add its tools, its permissions and its owner. An agent that is not in the inventory is not governed.

2. Named accountability, including for oversight. GOVERN 2.1 asks for documented roles, responsibilities and lines of communication; GOVERN 2.3 puts responsibility for decisions about AI risk on executive leadership; GOVERN 3.2 asks for policies that define roles for human-AI configurations and oversight. Every agent needs a named business owner, and every approval step a named reviewer.

3. A plan for retirement and for provider failure. GOVERN 1.7 covers decommissioning AI systems safely. GOVERN 6.2 asks for contingency processes for failures or incidents in third-party data or AI systems deemed high-risk, which for most agents includes the foundation model they run on.

MAP: what each agent is for and what it may touch

4. A written mandate. MAP 1.1 asks that intended purposes and deployment settings be understood and documented; MAP 3.3 that the targeted application scope be specified. For an agent, write down the tasks, the tools, the actions it may take and the ones it may not. That document is what every later check is measured against.

5. Oversight defined per action. MAP 2.2 asks for documentation of the system's knowledge limits and of how humans will oversee its output; MAP 3.5 for defined processes for human oversight. Decide which actions need approval before they run and which are reviewed afterwards, and check that approvals still involve real review.

6. Every component mapped, including third-party tools. MAP 4.1 and 4.2 ask that risks and internal controls be mapped for every component, including third-party software and data. For an agent, that means every tool and tool server it can call, because a tool's description can carry instructions of its own.

MEASURE: how to show an AI agent behaves as intended

7. Independent evaluation. MEASURE 1.3 asks for regular assessments involving internal experts who did not serve as front-line developers of the system, or independent assessors. The team that built the agent should not be the only one grading it.

8. Monitoring of behaviour in production. MEASURE 2.4 asks that the functionality and behaviour of the system be monitored in production; MEASURE 2.3 that performance be demonstrated for conditions similar to deployment. For an agent, monitor what it does, action by action, against its mandate, not just whether it is up. Pre-launch testing is not enough on its own: a red-team exercise is a snapshot.

9. Tested safe failure. MEASURE 2.6 asks that the system be shown to fail safely, "particularly if made to operate beyond its knowledge limits"; MEASURE 2.7 that security and resilience be evaluated. Test what the agent does when a tool is down, a check times out or a request falls outside its mandate, and decide per action whether it fails open or closed.

MANAGE: what happens when an AI agent misbehaves

10. A go/no-go decision per release. MANAGE 1.1 asks for a determination of whether the system achieves its intended purposes and whether its deployment should proceed. For an agent, make that a release gate with written criteria, and re-run it whenever the prompt, the tools or the underlying model change.

11. A stop path with owners. MANAGE 2.4 asks for mechanisms, with responsibilities assigned and understood, to "supersede, disengage, or deactivate" AI systems whose performance or outcomes are inconsistent with intended use. The Generative AI Profile adds that the criteria for deactivation should be set and regularly reviewed (MG-2.4-004). The mechanism itself should sit outside the agent it stops.

12. Post-deployment monitoring, override and incident response. MANAGE 4.1 asks for post-deployment monitoring plans that cover appeal and override, decommissioning, incident response, recovery and change management; MANAGE 4.3 that incidents and errors be communicated, tracked and responded to. An agent incident needs the same runbook as any production incident, plus a record of what the agent did and why.

A one-page NIST AI RMF summary for AI agents

FunctionAgent controlsSubcategories
GOVERNAgent inventory; named owners and reviewers; retirement and provider-failure plansGOVERN 1.6, 1.7, 2.1, 2.3, 3.2, 6.2
MAPWritten mandate; oversight per action; every tool mappedMAP 1.1, 2.2, 3.3, 3.5, 4.1, 4.2
MEASUREIndependent evaluation; production monitoring; tested safe failureMEASURE 1.3, 2.3, 2.4, 2.6, 2.7
MANAGERelease gate; stop path with owners; monitoring and incident responseMANAGE 1.1, 2.4, 4.1, 4.3

The order matters less than the links between them. A mandate written under MAP is what MEASURE tests against, and a failed measurement only means something if MANAGE gives someone the authority to act on it.

Where Syntrox fits

MEASURE 1.3 and 2.4 ask for evaluation that is independent of the builders and carries on in production. Syntrox sits outside the agent. Nucleus returns pass, fail or uncertain with a reason that quotes the fragment that drove the verdict, and returns uncertain rather than guessing when a grounded check has nothing to check against. Per evaluated trace, Syntrox records the score, confidence and judge version, so the measurement itself is traceable.