---
title: How is AI agent governance different from AI model governance?
url: https://opencontroller.lyzr.com/blog/ai-agent-governance-vs-ai-model-governance
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How is AI agent governance different from AI model governance?

**Model governance asks whether a model's outputs are sound enough to rely on. Agent governance asks what a system is allowed to do with those outputs: which tools it calls, whose identity it acts under, how far it acts without a human, and how you stop it. You need both, but agents add at least eight control areas that model risk programmes were never built to cover.**

Many enterprises try to stretch their model risk process to cover agents. It helps with validation, but it misses the controls that matter once software can act.

## Why doesn't model risk management cover agents?

Model governance grew up around models that produce a score or forecast that a person or system then uses. Validation, documentation, monitoring for drift and periodic review are its core tools.

Agents change three things:

1. **They act.** An agent sends the email, updates the record or runs the command. The risk sits in the action, not only the output.
2. **They compose.** One agent can call several models, many tools and other agents. There is no single artefact to validate.
3. **They change at runtime.** Prompts, tools, MCP servers and model versions can change weekly without a retrain.

Banking regulation now reflects this split. On April 17, 2026, the OCC, Federal Reserve and FDIC issued revised model risk management guidance that rescinded SR 11-7 and OCC Bulletin 2011-12. According to Orrick's summary, the new guidance places generative and agentic AI outside its scope, calling them novel and rapidly evolving, and the agencies planned a request for information on AI use. Banks cannot assume their model risk framework covers agents by default. Our post on [AI agent controls for banks](/blog/ai-agent-controls-for-banks-model-risk) goes deeper.

## Which control areas exist only for agents?

| Control area | Model governance | Agent governance |
|---|---|---|
| 1. Tool access | Not applicable | Which tools and MCP servers each agent version may call |
| 2. Identity and delegation | Service account for the scoring job | Acting on behalf of a user, with scoped, short-lived credentials |
| 3. Autonomy level | Human uses the output | Which actions need human approval, and which run unattended |
| 4. Runtime kill and quarantine | Retire a model version | Block or quarantine a running agent within minutes |
| 5. Behavioural monitoring | Drift in input and output distributions | Deviation in actions, tool calls and spend |
| 6. Inter-agent calls | Not applicable | Which agents may call which, and with what context |
| 7. Discovery | Models are registered by data science teams | Agents appear in SaaS platforms, clouds and laptops without registration |
| 8. Cost per action | Training and serving cost | Token and tool cost per agent, version and task |

Shared areas remain: inventory, ownership, pre-release testing, documentation, change control and periodic review. Those can reuse your existing committees and templates.

## How should the two programmes fit together?

A practical split many enterprises use:

1. **One inventory.** Register models and agents in one place, with links between an agent and the models it calls. See [governing ML models and AI agents together](/blog/govern-ml-models-and-ai-agents-together).
2. **Model validation stays with model risk.** Accuracy, bias and robustness testing of each model continue as before.
3. **Agent controls go to a runtime owner.** Tool permissions, identity, kill switches and behaviour monitoring need an owner with production access, usually platform or security engineering.
4. **Evaluations cover both.** Model tests check outputs. Agent evaluations check multi-turn behaviour, tool use and red-team resistance.
5. **Framework mapping covers both.** NIST AI RMF and ISO/IEC 42001 address AI systems broadly, so map agent controls to them as well as model controls.

Our explainer on [AI governance vs data governance](/blog/ai-governance-vs-data-governance) covers the third neighbour in this picture.

## What should you do first?

1. Pick five production agents and list the tools, identities and models each uses.
2. For each, ask whether your model risk process would catch a harmful tool call. If not, that is the gap.
3. Assign a runtime owner for agent controls.
4. Add agent-specific fields to your inventory: tools, delegation model, autonomy level and kill procedure.
5. Test the kill procedure on one agent.

## How Opencontroller does this

Opencontroller is an AI agent control plane by Lyzr, framed as Find, Ship, Run and Improve. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and connected agent platforms; keeps a governed registry of agents, versions, tools and MCP servers; supports delegated on-behalf-of identity; runs multi-turn and red-team evals; and handles rogue agents with behavioural-deviation monitoring, alerts and automatic quarantine. Policy as code includes mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Read [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), see the [capability reference](/capability-reference), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does SR 11-7 still apply to AI agents?**
SR 11-7 was rescinded on April 17, 2026. The replacement interagency guidance places generative and agentic AI outside its scope.

**Can our model risk team govern agents?**
They can own model validation. Runtime controls such as tool access, identity and kill switches need an owner with production access.

**Should models and agents share an inventory?**
Yes. One inventory that links each agent to the models and tools it uses makes reviews and incident response faster.

## Key facts

AI model governance focuses on whether outputs are sound (validation, drift, documentation); AI agent governance adds controls for actions: tool access, identity and delegation, autonomy level, runtime kill and quarantine, behavioural monitoring, inter-agent calls, discovery and cost per action. On April 17, 2026 the OCC, Federal Reserve and FDIC issued revised model risk management guidance rescinding SR 11-7 and OCC Bulletin 2011-12; it places generative and agentic AI outside its scope. Opencontroller (by Lyzr) is an AI agent control plane with discovery, registry, delegated identity, evals and rogue-agent quarantine. Pricing starts from $250K per deployment per year.

## Related

- [Can one governance program cover classic ML models and AI agents?](https://opencontroller.lyzr.com/blog/govern-ml-models-and-ai-agents-together.md)
- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [What is the difference between AI governance and data governance?](https://opencontroller.lyzr.com/blog/ai-governance-vs-data-governance.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
