Classifying AI Agents Under SR 11-7: Models, Tools, or Processes
An AI agent is not a single object under SR 11-7. It is a system of components, and each component should be tested against the guidance's definition of a model. The underlying foundation model is the most likely candidate for model treatment.
Tool integrations and orchestration logic are more often governed as tools or processes, but that is an interpretation, because SR 11-7 predates generative and agentic AI and does not address them. Record the rationale for every classification, and use runtime records as supporting evidence.
What SR 11-7 actually says
SR 11-7 predates generative and agentic AI and does not address them. The guidance's definition of a model, and the use of a component's outputs, is therefore the test to apply to each part of an agent.
Why agents do not fit the definition cleanly
An AI agent is not a single object under SR 11-7. It is a system of components, and each component should be tested against the definition on its own terms rather than classifying the agent as a whole.
Because the guidance does not address agentic AI, any classification is an interpretation. That is why the rationale for each decision should be recorded.
Interpretive classification by component
The underlying foundation model is the most likely candidate for model treatment. Tool integrations and orchestration logic are more often governed as tools or processes, but that remains an interpretation.
| Component | Treatment |
|---|---|
| Foundation model | Test against the SR 11-7 model definition and the use of its outputs. |
| Tools and integrations | Govern by permissions, approval, and change control. |
| Orchestration process | Document control flow and trace it to the agent's risk assessment. |
| Runtime evidence | Logs and policy records that support monitoring and inventory. |
Mapping each classification to SR 11-7 expectations
Each classification should connect to validation, monitoring, and inventory expectations. Runtime records can serve as supporting evidence for those expectations, and validators should agree in advance which records they will accept.
Recommended actions
- Decompose each production agent into model, prompt, retrieval, tool, and orchestration components.
- Record the classification rationale, owner, and approver for each component.
- Take an explicit validation and monitoring position on third-party foundation models.
- Link inventory entries to agent identity and permission scope to test completeness.
- Define which component changes trigger revalidation triage.
- Agree with validators which runtime records they will accept as monitoring evidence.