Use the vendor's admin controls, then verify. Team and Enterprise owners can disable Claude Code routines for all members, admins control Cowork access with a toggle, Claude Code's cron scheduler can be disabled with an environment variable, and ChatGPT admins can control event triggered tasks with a permission.
Controls found in vendor documentation
Read on 19 October 2026. Features change, so check your own admin console.
| Control | Where | Effect |
|---|---|---|
| Routines toggle | Claude admin settings for Claude Code, Team and Enterprise owners | Existing routines stop running and members cannot create new ones |
| CLAUDE_CODE_DISABLE_CRON set to 1 | Environment of the machine running Claude Code | Cron tools and loops become unavailable and scheduled tasks stop firing |
| Cowork admin toggle | Team and Enterprise admin settings | Controls access to Cowork, which includes scheduled tasks |
| Allow event triggered scheduled tasks | ChatGPT workspace admin permission | Controls tasks that start from Gmail, Slack or GitHub events |
| Cloud sessions policy | Organization policy for Claude Code | Routines require cloud sessions, so disabling them also blocks routines |
The routines page states that the routines toggle is a server side setting that cannot be overridden from local configuration. The environment variable, by contrast, is local, so it needs a way to be enforced on managed devices, such as your device management tool.
What these switches do not cover
- Personal accounts used outside the organization's plan. If staff use a free or personal login, your admin console does not see it.
- Tools from other vendors, which have their own controls.
- Scheduled workflows in your repositories, which are controlled through repository and organization settings in your code host.
- Local tasks on machines where no policy is applied.
Pair the switches with detection. Network and device inventories can show which assistant apps are installed, and the shadow automation guide covers discovery.
Switch off, or narrow?
A blanket off is the cleanest response to an incident. For day to day governance, narrower choices are often better:
- Allow scheduled tasks for read-only use cases first.
- Require a sponsor and registry entry before enabling for a team.
- Limit connectors and environments to approved ones. Anthropic notes an Owner can manage shared cloud environments from admin settings.
- Enforce managed settings for coding agents, as in enforcing team settings.
Testing that the switch worked
After flipping a control, do not assume. Pick a test task and confirm it stops. For routines, the page says existing routines stop running when disabled. Create a harmless task in a test account first, flip the switch, and watch that its next run does not occur. Record the date and result as evidence.
Communicating the change
Tell staff what changed and why, point to the approved alternative, and list the exception process. Quiet switch offs push people toward personal accounts.
Worked example: pausing automation during an incident
At 14:00 a security team learns that a connector used by several automations may have been compromised. They need to stop scheduled agents from using it within minutes, without breaking everything else. Here is a staged response using the controls described above.
Stage one is containment of the credential. They revoke the connector grant at the source service. That stops anything using it from working, whatever the assistants are doing.
Stage two is stopping the schedulers that touch it. Owners disable the routines that include the connector, or an owner uses the organization wide routines toggle if the blast radius is wide and time matters. Existing routines stop running, per the documentation. For local tooling, device management pushes the environment variable that disables Claude Code's cron scheduler to managed machines. For event triggered tasks in other products, an admin removes the permission that allows them.
Stage three is verification. They create no new tasks but check the run histories for runs after the cutoff. A run after the cutoff would show a control that did not work.
Stage four is restoration, one team at a time, with the connector reauthorized at a narrower scope.
Choosing the right granularity
| Need | Control |
|---|---|
| Stop everything in one product for everyone | Organization toggle, such as the routines setting |
| Stop local scheduling on managed machines | Environment variable pushed by device management |
| Stop one team | Remove their access or the specific tasks |
| Stop one task | Pause it, then revoke what it used |
| Stop event triggers only | The admin permission for event triggered tasks, where offered |
Pitfalls
- Assuming the toggle covers other products. Each vendor has its own controls. Keep a list of products and what each switch does.
- Forgetting the personal account route. If staff move to a personal login after you disable the managed one, you have moved the risk out of sight.
- No test. Run the switch in a sandbox first so you know the effect and the latency.
- No communication. Staff who do not know why a tool stopped will look for workarounds.
- No restore plan. Decide in advance who may turn things back on and under what conditions.
Record it
Keep a short runbook: which controls exist, who may operate them, how to verify, and the date each was last tested. Link it from the incident plan and from the AI agent kill switch guide.
Try it yourself
Click through it at your own pace, or open it full screen.
Look at how a server connection is defined, and note where you would cut it off in the staged response above.
Quick check
If an owner disables routines, do existing ones keep running?
Per Anthropic's routines page, no. Existing routines stop running and members cannot create new ones.
Does an admin toggle cover personal accounts?
No. It applies to the organization's plan, so decide how to handle personal logins in policy.
Who should be allowed to flip an organization wide switch?
A small named group, such as security on call and the platform owner, with a written trigger and a record of each use.
Sources
- Automate work with routines (Claude Code docs)
- Run prompts on a schedule (Claude Code docs)
- Schedule recurring tasks in Claude Cowork (Help Center)
- Scheduled tasks and automations (OpenAI Codex docs)
Last reviewed: 19 October 2026.
