---
title: When does a change to an AI agent need re-approval?
url: https://opencontroller.lyzr.com/blog/when-does-an-ai-agent-change-need-re-approval
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# When does a change to an AI agent need re-approval?

**An AI agent needs re-approval when a change can alter what it is able to do, what data it touches or how much it decides on its own: a new model, a new tool or MCP server, broader permissions, a new data source or higher autonomy. Prompt wording tweaks, bug fixes and configuration patches can usually ship through normal review if evaluations still pass. Classify every change up front so teams know which path they are on before they ship.**

Most governance programmes approve an agent once, at launch. Agents then change weekly. The approval drifts out of date unless you define which changes reopen it.

## Why can't you treat agent changes like normal code changes?

Because an agent's behaviour depends on more than its code. Four things can change it without a line of application code moving:

- The model: a provider version bump, or a swap to a cheaper model.
- The system prompt and instructions.
- The tools, MCP servers and skills it can call.
- The data and permissions behind those tools.

A model swap can change refusal behaviour, tool-calling accuracy and cost in one step. A new MCP server can give the agent write access to a system nobody reviewed. Code review alone catches neither.

## Which changes need re-approval?

Use three change classes. The class decides the path, the evidence and who signs.

| Class | Examples | Path | Evidence required |
|---|---|---|---|
| Material | New or different model family; new tool, MCP server or skill; new write permission; new data source with personal or regulated data; higher autonomy; new user population or external exposure; new jurisdiction | Full re-approval by the risk owner, same as launch | Updated registration record and risk tier, eval results against the previous version, red-team results for new tools, data-flow note |
| Minor | Prompt wording changes; same-family model patch; retry, timeout or caching settings; bug fixes; UI text | Normal change review by the agent owner and a peer | Eval suite pass at or above the approved baseline, prompt and config diff, version record |
| Emergency | Blocking a tool, rolling back a model, tightening a guardrail, quarantining the agent | Act first, review within an agreed window, for example two business days | Incident record: who acted, what changed, follow-up decision |

Two rules keep the classes honest:

1. **Tightening never waits; loosening always does.** Removing a tool, narrowing a scope or lowering autonomy can ship immediately. Adding any of them is material.
2. **When in doubt, classify up.** A "minor" change that touches a high-risk agent's tool permissions is material.

## What should trigger the check automatically?

Self-classification fails at scale. Wire triggers into the places changes actually happen:

1. **Registry diff.** Compare each new agent version with the last approved one: model ID, tool list, MCP servers, scopes, data sources and autonomy setting. Any change in those fields flags a material review.
2. **CI/CD gate.** Block deployment of a version whose eval scores fall below the approved baseline.
3. **Gateway traffic.** Alert when an agent starts calling a model or endpoint that is not in its approved record.
4. **Identity changes.** Alert when an agent's service principal or role gains a permission.
5. **Provider notices.** Track model deprecation dates. A forced model migration is a material change on someone else's schedule.

## How does the risk tier change the rules?

Re-approval should scale with risk. A low-risk internal summariser can treat a model upgrade as minor if evals pass. A customer-facing agent that issues refunds should treat any model change as material. Link change classes to your [risk tiers](/blog/how-to-risk-tier-ai-agents) and write the mapping into policy, not into people's heads.

A practical default:

- **Tier 1 (high risk):** every model change, and every prompt change that touches tool use or disclosure, is material.
- **Tier 2 (medium):** model family changes and any new tool are material.
- **Tier 3 (low):** only new tools, new data classes and new write permissions are material.

## What do auditors expect to see?

Frameworks ask for controlled change, not one specific process. ISO/IEC 42001 requires organisations to control planned changes and review the consequences of unintended ones as part of operational planning and control (clause 8.1). The NIST AI RMF MANAGE function covers how AI risks are tracked and responded to after deployment. The evidence for both looks the same in practice:

- A version history for each agent showing who approved what, and when.
- Eval results attached to each approved version.
- Records of emergency actions and their follow-up review.

Tie each record to your [production readiness gate](/blog/production-readiness-gate-for-ai-agents) so launch approvals and change approvals use the same criteria, and to the [agent lifecycle](/blog/ai-agent-lifecycle-stages) so retired versions stay traceable.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so each new version can be compared with the approved one. Multi-turn and red-team evals, and LLM judges scoring production traffic, give before-and-after evidence. The LLM Gateway attributes cost per agent and per version, which shows when a model swap changes spend. Policy as code lets you encode which fields count as material. Emergency actions such as Block and Quarantine act through the connected platform. Opencontroller does not ship human-approval hold workflows in the gateway; approvals run in your existing change process. Read the [control plane overview](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does a provider model upgrade need re-approval?**
It depends on the risk tier. For high-risk agents, treat any model change as material. For low-risk agents, a same-family update can ship as minor if the eval suite passes at or above baseline.

**Is a system prompt change a material change?**
Usually not, unless it changes which tools the agent may use, what it may disclose or how autonomously it acts. Those prompt changes are material.

**Who should approve a material change to an AI agent?**
The role that approved launch, typically the risk owner for that agent's tier, with the agent owner supplying evidence. The author of the change should not approve it.

## Key facts

Material AI agent changes needing re-approval: new model family, new tool/MCP server/skill, new write permission, new sensitive data source, higher autonomy, new users or jurisdiction. Minor changes (prompt wording, same-family patch, config) need eval pass at baseline. Emergency tightening acts first, reviews later. ISO/IEC 42001 clause 8.1 covers control of planned changes. Opencontroller provides a versioned registry, evals, LLM judges, per-version cost attribution and Block/Quarantine; it does not ship gateway human-approval hold workflows. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [What are the stages of the AI agent lifecycle?](https://opencontroller.lyzr.com/blog/ai-agent-lifecycle-stages.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
