---
title: Should AI agents follow change freezes?
url: https://opencontroller.lyzr.com/blog/change-freezes-for-ai-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Should AI agents follow change freezes?

**Yes. A prompt edit, a model swap, a new tool or a policy change alters what an agent does, so each one belongs inside your change freeze. The exceptions are safety actions that reduce risk, such as blocking or quarantining an agent, which should stay allowed during a freeze.**

Many teams freeze code before peak trading or financial close, then discover agents changed behaviour mid-freeze because someone edited a prompt in a console. This post sets out what counts as a change and how to scope a freeze. It is part of our [runtime control](/blog/how-to-stop-a-rogue-ai-agent) series.

## What counts as a change for an AI agent?

Code is only one part of an agent. Behaviour also depends on configuration that often lives outside source control.

- **Prompts and instructions.** System prompts, agent instructions files, few-shot examples.
- **Model.** Provider, model name and version, and parameters such as temperature or reasoning effort.
- **Tools and MCP servers.** Adding, removing or changing a tool's scope or schema.
- **Knowledge sources.** New documents or indexes the agent retrieves from.
- **Policies and guardrails.** Thresholds, block rules, redaction settings.
- **Routing.** Gateway rules that send the agent to a different model or fallback.
- **Permissions.** New roles, scopes or data access.

Change management frameworks already cover this in principle. SOC 2's common criteria include change management (CC8.1), and NIST SP 800-53 CM-3 covers configuration change control. Agents just make the "configuration" part much larger.

## Which changes should a freeze block?

| Change type | Freeze status | Reason |
|---|---|---|
| Prompt or instruction edit | Blocked | Directly changes behaviour |
| Model or model version change | Blocked | Output quality and cost shift |
| New tool or MCP server | Blocked | New actions and data paths |
| Expanded permissions or scopes | Blocked | Larger blast radius |
| New knowledge source | Blocked | New content in answers |
| Relaxing a guardrail or policy | Blocked | Removes a control |
| Tightening a guardrail or policy | Allowed with record | Reduces risk |
| Blocking or quarantining an agent | Allowed | Safety action |
| Revoking credentials | Allowed | Safety action |
| Gateway fallback to a pinned backup model | Allowed if pre-approved | Resilience, defined in advance |

The principle: changes that reduce risk are allowed; changes that add capability or remove controls wait.

## What about model changes you do not control?

Providers retire models on their own schedules. Anthropic, OpenAI, Google and others publish deprecation dates for older models, and an alias that points to "latest" can move without your release.

Before a freeze:

1. **[Pin model versions](/blog/pin-model-versions-for-ai-agents)** instead of using floating aliases.
2. **Check provider deprecation pages** for any retirement inside the freeze window.
3. **Pre-approve a fallback model** and test it, so a forced switch is a known change.
4. **Freeze gateway routing rules** except for pre-approved failover.

## How do you enforce a freeze on agents?

A freeze written in an email is not enforced. Make it technical.

1. **Inventory first.** You can only freeze agents you know about. Include agents built in SaaS platforms, not only your code repositories.
2. **Put configuration in version control or a registry** so every change has a record and an author.
3. **Restrict console edit rights** during the window. Many agent platforms allow prompt edits in the browser.
4. **Use policy as code** to reject deployments of new agent versions during the freeze window.
5. **Detect drift.** Compare running configuration with the approved version daily during the freeze and alert on differences.
6. **Watch behaviour.** A [behavioural drift alert](/blog/detect-behavioural-drift-in-ai-agents) can catch changes that bypassed the process.

## How should emergency changes work?

Some changes cannot wait. Define the exception path before the freeze:

- Named approvers (usually the agent owner plus a change authority).
- A ticket that records the reason, the change and the rollback plan.
- A [regression test run](/blog/regression-test-ai-agent-after-prompt-change) before the change goes live.
- A review after the freeze ends.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, and auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, with connectors to third-party agent platforms. Policy as code applies rules across the estate, behavioural-deviation monitoring raises alerts, and per-agent Block and Quarantine controls remain available as safety actions during a freeze. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is a prompt change really a production change?**
Yes. A prompt edit can change what an agent says, which tools it calls and what data it shares. Treat it like a code change with an author, review and record.

**Can we still block a misbehaving agent during a freeze?**
Yes. Blocking, quarantining and revoking credentials reduce risk and should always be allowed, with a record.

**What if a provider retires our model during the freeze?**
Plan for it before the window: check deprecation dates, pin versions, and pre-approve and test a fallback model so the switch is a known change.

## Key facts

AI agent change freezes should cover prompts, model versions, tools and MCP servers, knowledge sources, permissions, routing and relaxed policies. Risk-reducing actions (blocking, quarantining, revoking credentials, tightening policy) stay allowed. SOC 2 CC8.1 and NIST SP 800-53 CM-3 cover change management. Before a freeze, pin model versions, check provider deprecation dates and pre-approve a fallback model.

## Related

- [How do you stop a model update from silently changing agent behaviour?](https://opencontroller.lyzr.com/blog/pin-model-versions-for-ai-agents.md)
- [How do you roll out and roll back a new AI agent version safely?](https://opencontroller.lyzr.com/blog/canary-releases-and-rollback-for-ai-agents.md)
- [How do you regression-test an AI agent after changing its prompt?](https://opencontroller.lyzr.com/blog/regression-test-ai-agent-after-prompt-change.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
