> ## Documentation Index
> Fetch the complete documentation index at: https://docs.visiqlabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Kill Switch

> Block a harnessed agent's activity for a set number of days, or until someone releases it.

## Overview

The Kill Switch stops an agent you have harnessed with VisIQ from acting. Open the agent's page, press **Kill Switch**, choose how long, and confirm. VisIQ then refuses the agent's tool calls and data access, whatever mode the agent is in, usually within 30 seconds. The agent's process keeps running, and its record and history stay.

A blocked agent keeps reporting to VisIQ. Everything it tries is refused and shows in its activity, so you can see what it attempted while it was blocked.

Pressing the Kill Switch also refuses the agent's requests that are waiting on a person in the [approval queue](https://docs.visiqlabs.com/rules/action/hitl.md). While the agent is blocked, none of its requests can be approved, from the queue, Slack or an email link.

Work the agent hands to another agent is blocked too. Actions taken under a blocked agent's delegation are refused, and a blocked agent cannot accept work delegated to it.

Every agent whose Kill Switch is on is listed on **Control AI → Agent Jail**, with when it was blocked, when the block ends and who started it. Release an agent from that page or from the agent's own page, and it returns to the mode it had before, usually within 30 seconds. A running process carries on without a restart. A process that stopped while it was blocked needs starting again.

The Kill Switch is only offered for harnessed agents. VisIQ is not in the call path of an agent it only discovered, so there is nothing to refuse.

***

## How long it lasts

| Time frame | When it ends |
| - | - |
| **Indefinitely** (the default) | When someone releases the agent |
| **A custom number of days**, 1 to 365 | Automatically, within a few minutes of the last day ending. Until then the agent stays blocked, and Agent Jail shows the release as due |

To change a time frame, release the agent and start the Kill Switch again.

Two things turn a block indefinite whatever was chosen, and neither can be undone except by a person releasing the agent:

* **Containing the agent on a malicious verdict.** Any time limit ends there, and Agent Jail names Detection as the source. Deal with the alert before releasing the agent.
* **A copy of a blocked agent.** When a new agent ID registers from the same key, host and user as a blocked agent, VisIQ holds it with no end date. Agent Jail shows it as held automatically.

***

## What it does not stop

The Kill Switch works where VisIQ gates an agent, so it has the same limits as a rule:

* **Anything already running.** A tool call in progress finishes. The process is not stopped.
* **Anything the harness does not gate.** Model calls, plain text replies and tools that are not wrapped by the harness are not refused.
* **Claude Cowork sessions in the VM sandbox.** VisIQ monitors that surface and cannot block it. Host-local Cowork sessions are blocked as usual.
* **An agent that cannot reach VisIQ.** It keeps the rules it last read, and one starting with no rules follows its fail mode.
* **A different identity.** A new agent ID registered from a different key, host or user is a new agent. When a blocked agent has no key on record, only its host and user are compared.

How soon the block lands depends on how the agent reads its rules:

| Harness | When the block lands |
| - | - |
| TypeScript SDK, OpenClaw | Within 5 seconds |
| Claude Code, Codex, Cline and other CLI harnesses | At the next tool call, within 30 seconds |
| Python SDK | At the next gated call |
| Java, Go, Ruby and Rust SDKs, using the `Governor` | Within 5 seconds |
| Java, Go, Ruby and Rust passing in a bundle you fetch yourself | When you next fetch [the rule bundle](https://docs.visiqlabs.com/rules/unified/api-reference.md). Fetch it at least every few seconds |

To stop a process, stop it on its host. Revoking its API key does not stop it. A process whose key is revoked keeps the rules it last read, so if you revoke the key before the block lands, the block never reaches it. If a key has leaked, wait for the block to land, then revoke the key. A process that later starts from a clean disk has no rules and follows its fail mode.

Deleting a blocked agent ends the block: a process still holding a valid API key registers again on its next call and can act. It takes the `allow_agents:delete` permission, even for the agent's owner.

A blocked agent cannot be moved to another account. Release it first.

***

## Who can do what

| Action | Permission |
| - | - |
| Start the Kill Switch | `allow_agents:update` |
| Release an agent | `allow_agents:update`, signed in as a person |
| Delete an agent that is blocked | `allow_agents:delete`, signed in as a person |
| See Agent Jail | `allow_agents:view` |

An API key can start the Kill Switch, but it cannot release a blocked agent or delete one. Only a signed-in person can, so an agent that gets hold of a key cannot set itself free.

Starting the Kill Switch is written to your organization's audit log as `ALLOW_AGENT_KILL_SWITCH_ON`, releasing an agent or deleting a blocked one as `ALLOW_AGENT_KILL_SWITCH_OFF`, and an automatic release as `ALLOW_AGENT_KILL_SWITCH_EXPIRED`. A block from a malicious verdict is recorded on its alert, and an automatic hold shows in Agent Jail.

***

## From MCP

The [MCP server](https://docs.visiqlabs.com/automation/mcp.md) exposes the same choice on `update_agent`: `shutdown: true` with an optional `shutdown_days`, leaving it out for an indefinite block, and `shutdown: false` to release. `list_agents` takes `jailed: "true"` to list blocked agents.

A blocked agent is denied with `rule_code` `AGENT_SHUTDOWN` on [`/allow/evaluate`](https://docs.visiqlabs.com/rules/action/api-reference).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.