Guide 54 of 61 · Running a fleet with Central Lite
Stop an agent task on any machine
Kill or suspend one agent task from the Central Lite host, check that the node applied it, and revoke a machine you no longer trust.
Built, not yet claimed preview: this guide describes a package that has not passed its release gate.
Find the machine in Central Lite's inventory and the task's id on the machine, queue a kill (or a suspend you can resume) from the Central Lite host, and check that the node applied it: the task is denied every decision and its processes are killed or stopped. Then, if the machine itself is no longer trusted, revoke it.
1. Find the machine
Every kill names a node by its host id, a UUID. In the console, open Inventory: the Nodes table, first on the page, lists each node with its Host, State (enrolled or revoked; a node that left shows as revoked) and Last report, and its Host id from the table's Columns menu (the first 8 characters; the whole id on hover). On the Central Lite host, garm-central node list prints host id, hostname, state (enrolled, revoked or left) and last report.
garm-central node show HOST shows one node in full, including open cmds, the commands Central Lite sent that the node has not acknowledged yet. A node that left Central Lite cannot be sent a kill.
garm-central node listgarm-central node show HOSTChecked by the walk: this command ran as written on a Central Lite host.
See also
Enrol developer machinesEnrol a machine in Central Lite with a one-time token that pins Central Lite's key, see it report in the inventory, and let its users enrol presence keys.
2. Read the task's id on the machine
Central Lite does not list a node's live tasks, so the task id comes from the machine. There, agentgarm status --json lists each live task under tasks with its full task_id, its agent and its assurance; run as root it shows every user's tasks, and run as a user only that user's. The plain output shows only the last twelve characters of each id, and a kill needs the whole UUID: Central Lite refuses anything else.
sudo agentgarm status --jsonChecked by the walk: this command ran as written on a Linux node enrolled in Central Lite.
See also
Trace what one agent task didFind an agent's task in the dashboard, read every decision its policy made with the rule behind it, and follow the run from first decision to last.
3. Queue the kill
On the Central Lite host, garm-central node kill queues a kill of one task on one node and prints kill with the command's id, then queued for task with the task and the host. The reason goes to the node and into both audit logs; without --reason it is killed by an admin. The action is chained in Central Lite's audit log as KILL_QUEUED, with the operating-system account that ran it. The console has no single-task kill: it is a host command only.
A kill targets one task. To stop every task on a machine, or agents across many, request a kill order instead, which several approvers sign and each node checks (Stop agents across a fleet). Central Lite never sends a kill of every task on a node (all_tasks), and a node refuses one if it arrives. While serve runs, the kill goes to a connected node within a second. It is sent again until the node acknowledges it, so a node that is offline gets it when it reconnects, if the task is still running.
garm-central node kill HOST --task TASK_ID --reason "runaway agent"4. Suspend a task instead, and resume it
When you want the task stopped but not ended, garm-central node suspend queues a suspend of one task on one node and prints suspend … queued for task … on …; without --reason the reason is suspended by an admin. The node stops the task under a hold of its own, command: and the task id, which agentgarm status lists under the task. garm-central node resume removes that hold, and only that one: a budget breach's or a kill order's hold stays, and the task runs again once none is left. Root on the machine can also remove it with sudo agentgarm task resume.
Central Lite refuses a suspend for a node whose agentgarmd is too old to know one, since such a node would end the task instead. A suspend counts toward the node's 10 commands an hour, as a kill does. Resume is not built on Windows yet: a task suspended there stays suspended until it is killed.
garm-central node suspend HOST --task TASK_ID --reason "paused for review"
garm-central node resume HOST --task TASK_IDSee also
Find out why your agent stoppedTell whether your agent was suspended, killed, locked out by a kill order or stopped by its budget, get your work back, and find out who can lift it.
5. Know what the node and the user see
The node marks the task suspended first, so every later decision for it is denied with this task is suspended. Then it kills the task's processes: where the task has its own cgroup (Linux) or Job Object (Windows), every process in it is killed, subagents included; otherwise agentgarmd sends SIGKILL to the task's root process and its MCP servers' process groups, and on macOS to the processes Endpoint Security tracked in its tree. For the user, the agent stops mid-task. The mode is fail_closed, enforced by the cgroup or the signal, not by hooks. If the kill itself fails, the task stays suspended: every decision it asks for is denied, which in a hooks-only session is advisory.
Every kill, applied or refused, is chained in the machine's own audit log with the task and the reason; an applied kill is alerted on that task, a refusal as a host alert. A node refuses a task that is not live there, and more than 10 kills from Central Lite in an hour. A single-task kill, suspend or resume is not signed by anyone: the node takes it from the Central Lite it trusts over its pinned mutual-TLS channel, and the hourly cap bounds what that channel can do, so whoever controls Central Lite's host can queue one. A kill order is different: it carries its approvers' signatures, which every node checks itself.
6. Check that the kill landed
Once the node acknowledges the kill, garm-central node show HOST no longer counts it under open cmds. A refusal raises a command_refused alert in Central Lite, with the node's reason: for example, no live task has this id on this node, or more than 10 kills within an hour. garm-central alerts list shows unacknowledged alerts, newest first, and alerts ack acknowledges one, chained.
garm-central alerts listChecked by the walk: this command ran as written on a Central Lite host.
7. Revoke the machine if you no longer trust it
Revoking a node revokes every certificate it holds and issues the next revocation list naming them. Every Central Lite service refuses those certificates at once, and the node's control stream ends. garm-central node revoke prints how many certificates were revoked and the list's number. In the console, choose Revoke in the node's row: its sheet asks “Why revoke …? (chained)”, and “Revoke with a passkey” asks for a fresh passkey assertion for this one action.
Revoking stops no running task. Kill the task first: once the node is revoked it accepts no command from Central Lite, kills included, and uploads nothing until it is enrolled again, though it keeps every unsent event.
garm-central node revoke HOST --reason "machine compromised"