---
title: How do you tell users when an AI agent is shut down or quarantined?
url: https://opencontroller.lyzr.com/blog/notify-users-when-an-ai-agent-is-shut-down
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you tell users when an AI agent is shut down or quarantined?

**Trigger notifications from the quarantine or kill event itself, not from a person remembering to send an email. Notify four groups: the agent owner, the people and systems that use the agent, any agents that call it, and the incident channel. Give users a clear in-product message and a fallback route so the work keeps moving.**

When security stops an agent, the worst outcome for users is a silent failure. A chat window that times out, a workflow that hangs, or an API that returns a 500 looks like a bug. People retry, open tickets, or route around the control. A good shutdown notice turns a security action into a normal, understood event.

## Who needs to know when an agent is stopped?

Each audience needs a different message and a different channel.

| Audience | What they need to know | Channel | When |
|---|---|---|---|
| Agent owner | Why it was stopped, the policy or signal that fired, the evidence, how to request reinstatement | Direct alert plus ticket | Immediately |
| End users | The agent is unavailable, what to do instead, when to expect an update | In-product message, status page | On next request |
| Calling systems and agents | A machine-readable "unavailable" response with a reason code | API error or tool result | On next call |
| Security and operations | The event, the scope, the blast radius | Incident channel, SIEM, ITSM | Immediately |
| Business sponsor | Business impact and expected duration | Email or chat summary | Within the hour for high-impact agents |

The owner and security team get the full reason. End users usually do not. "Quarantined for suspected data exfiltration" in a customer-facing chat window creates more problems than it solves. Keep the user message neutral and the detail in the owner's ticket.

## What fallback should users get?

A shutdown notice without a next step just moves the frustration. Decide the fallback when the agent is registered, not during the incident. Record it as a field in the agent's inventory record.

Common fallback routes, in rough order of preference:

1. **Last approved version.** If the problem came with a new version, route traffic back to the previous version that passed review. This only works if versions are tracked and the old one is still deployable.
2. **Reduced mode.** Keep read-only or answer-only behaviour and disable the tools that take actions. Useful when the risk is in one tool, not the whole agent.
3. **Human queue.** Send requests to the team that did the work before the agent existed. Include the user's original request so they do not have to repeat it.
4. **Another approved agent.** Only if a second agent is already approved for the same task and data.
5. **No fallback, clear message.** For low-impact agents, a clear message and an expected update time is enough.

Pick the fallback by risk tier. A customer-facing agent that handles refunds needs a human queue. An internal summarisation agent can show a message and wait.

## What should the user-facing message say?

Keep it short, honest and actionable. A template that works for most internal agents:

> **[Agent name] is temporarily unavailable.**
> We paused this assistant while we check it. Your request has not been processed.
> **What you can do now:** [fallback, for example "submit the request through the service desk form" with a link].
> **Next update:** [time] on [status page link].

Three rules for the wording:

- **Say whether the request was processed.** Users need to know if a refund, email or update went out.
- **Do not blame the user** or imply their input caused the shutdown, even if it did.
- **Give a time for the next update,** not a time for the fix. You rarely know the fix time.

For calling systems, return a consistent error. Use a stable reason code such as `agent_quarantined` or `agent_disabled`, an HTTP 503 with a `Retry-After` header where a retry makes sense, and the incident or ticket ID. For an agent exposed as an MCP tool, return a tool result with `isError` set to true and a plain-language message, so the calling model can explain the outage instead of retrying in a loop.

## How do you wire notifications to the quarantine event?

Manual notification fails under pressure. Automate it from the event that stops the agent.

1. **Emit one event per state change.** Active to quarantined, quarantined to disabled, and back. Include agent ID, version, owner, policy or signal, severity, and timestamp.
2. **Look up the audiences from the registry.** Owner, business sponsor, consumer groups and dependent agents should be fields on the agent record. If they are missing, the notification has nowhere to go.
3. **Send to each channel from one rule.** Ticket to ITSM, message to the incident channel, alert to the owner, banner to the status page.
4. **Switch the front door.** The gateway or the platform returns the user message and reason code instead of a timeout.
5. **Notify on reinstatement too.** Users who were told the agent was paused should be told it is back.
6. **Log every notification** with the incident record. Auditors and post-incident reviews will ask who was told and when.

Test this in the same exercise where you [test your AI agent kill switch](/blog/test-your-ai-agent-kill-switch). A kill switch that works but leaves users staring at a spinner has only passed half the test.

## How does this fit the wider incident process?

Notification is one step in an [AI agent incident response runbook](/blog/ai-agent-incident-response-runbook). NIST SP 800-61 Rev. 3 frames incident response around communication with internal and external parties as part of handling, not as an afterthought. The same applies to agents. Decide in advance which incidents need customer or regulator communication, and who approves it.

For deciding when the system should pause an agent without waiting for a person, see [when to automatically quarantine an AI agent](/blog/when-to-automatically-quarantine-an-ai-agent). The full playbook is in [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).

## How Opencontroller does this

Opencontroller monitors agents for behavioural deviation, raises alerts, and can automatically quarantine or kill an agent under policies that apply across the estate. Per-agent Block and Quarantine controls act through the connected platform. The governed registry holds the agent, its versions and its owner, which is where notification routing starts. Policies are written as code, so the rule that quarantines an agent and the rule that alerts its owner live together. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should end users be told why an AI agent was shut down?**
Usually not in detail. Tell them the agent is paused, whether their request was processed, and what to do instead. Give the full reason to the owner and the security team.

**What should a calling application receive when an agent is quarantined?**
A consistent, machine-readable error: a stable reason code, an HTTP 503 with Retry-After where retrying makes sense, and an incident ID. Avoid timeouts, which trigger blind retries.

**Do you need to notify users when the agent comes back?**
Yes. Anyone told the agent was paused should hear it is restored, ideally through the same channel and with any change in behaviour noted.

## Key facts

When an AI agent is quarantined or shut down, trigger notifications automatically from the quarantine event. Notify the agent owner (full reason and evidence), end users (neutral in-product message saying whether their request was processed, a fallback and the next update time), calling systems and agents (stable reason code, HTTP 503 with Retry-After, or an MCP tool result with isError true), and the security/incident channel. Decide fallback routes at registration: last approved version, reduced mode, human queue, or another approved agent. Opencontroller by Lyzr provides behavioural-deviation monitoring, alerts, automatic quarantine and kill, and per-agent Block and Quarantine controls that act through the connected platform.

## Related

- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [What should an AI agent incident response runbook include?](https://opencontroller.lyzr.com/blog/ai-agent-incident-response-runbook.md)
- [How do you test that your AI agent kill switch actually works?](https://opencontroller.lyzr.com/blog/test-your-ai-agent-kill-switch.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
