One agent
Stop a single agent from its own page or from the fleet page.
Kill switch
One switch stops an agent's model calls, its tool calls and its approvals. It needs a reason, it has its own permission, and it stays on until a person lifts it.
Scope
Stop a single agent from its own page or from the fleet page.
Stop every agent that shares a tag, a kind or a blueprint, or a list you pick.
One switch stops them all.
Pause a single tool or a whole connected server, and the agents carry on without it.
Effect
The switch is checked before each model call, each step of a run, each new token and each approval, and on calls that arrive over MCP and A2A.
A model call or a tool call already in flight is dropped within seconds, and the run ends there.
Turning a switch on needs a reason, and the audit log records who did it and when.
Stopping agents is a separate permission, so the on-call team can hold it without being able to edit an agent.
A stopped agent stays stopped until a person lifts the switch. It does not come back by itself.
Set a detector on a signal, such as the share of calls the guardrails refuse, the error rate or a changed tool definition. It can report only, warn, or stop the agent.
Questions
New calls are refused at once, because the switch is checked before each one. A model call already in flight is dropped within seconds, and the run ends there.
Anyone you give the permission to. It is separate from the permission to build or edit agents.
Yes, for what passes through Liyagent. An agent that sends its model calls through the Liyagent gateway is refused there while its switch is on.
No. Only a person with the permission can lift a switch, and a stopped agent is refused a new token.
Related