Skip to the content

Find out why your agent stopped

Tell 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.

Built, not yet claimed preview: this guide describes a package that has not passed its release gate.

Read what agentgarm status says about your machine and your tasks, tell a suspend, a kill, a lockdown and a budget breach apart, know what keeps working meanwhile, get your work back with rollback, and find out who can lift what stopped you.

1. See what happened

Run agentgarm status. A kill order in force on this machine comes first: a lockdown as LOCKDOWN, “this machine is locked down: no governed agent starts or acts”, any other order as a kill order line, each followed by the order's id, its scope, who approved it, the custody of the org root it was checked against, its reason, and that it is lifted only by its lift, or by root on this machine. Under a task, SUSPENDED and since when means it is stopped until every hold listed below it is lifted.

The local dashboard's Overview opens with the kill orders in force on this machine, a lockdown first (“This machine is locked down”), each with who can lift it. Its Alerts list each hold placed or removed under Stopped tasks, and History marks a suspended run with a suspended badge, its holds in the badge's title. The menu bar (or the tray on Windows) also shows a notice when a kill order, a Central Lite admin or your budget suspends a task, and when the machine is locked down or the lockdown is lifted; it names who stopped the task, never the order's reason.

agentgarm status

Checked by the walk: this command ran as written on a no-root Linux install.

See also

Check that protection is on

See whether AgentGarm's services are up, what assurance level each agent session has and why, and what to do about a bypass alert.

2. Tell a suspend, a kill, a lockdown and a budget breach apart

Each hold under a suspended task says who placed it. order: and an id is a kill order approved by several people in Central Lite. command: and a task id is a single-task suspend a Central Lite admin sent. breach is your task's own budget or loop limit. airbag and sensitive_frontmost come from the computer-use airbag and the screen-capture monitor on a Mac. A suspended task keeps its processes, stopped, and every decision it asks for is denied with “this task is suspended”, or “a kill order is in force”.

A killed task is gone: its processes were ended, and agentgarm status shows it ENDED until it drops off the list. Under a lockdown, or a suspend or terminate order that covers you, agentgarm run refuses to start a new agent with agentgarmd's reason: “this machine is locked down by kill order …” or “kill order … refuses new tasks in its scope”, with the order's scope and reason.

See also

Set spend budgets and stop runaway loops

Give each agent task a cost budget and a token rate, see what the agent meets at the cap, and stop loops and runaway subagents with limits of their own.

3. Know what still works

An order stops only governed agents: tasks started with agentgarm run or agentgarm wrap, the subagents and MCP servers in them, and hooks-only sessions, whose every decision is denied. Your shell, your editor and any tool you run yourself keep working, and agentgarmd never signals your IDE. Ungoverned processes are untouched, and the record says so.

For a suspended task, its streamed model responses are ended and its network tunnels closed, so the agent sees an incomplete stream. How firmly it is held depends on the machine: on Linux with the task's own cgroup it is frozen (A); on macOS it is stopped by signals, which another process of yours can undo until the next sweep (B with the Endpoint Security client, C without); a hooks-only session or a no-root install is C.

4. Get your work back

Nothing is checkpointed at kill time. Rollback restores the checkpoint agentgarm run took when the task started, and only once the task has ended: it refuses a task that is still running, and a suspended task still is. After a terminate or a lockdown, run agentgarm rollback with the task's id (agentgarm task list --ended lists it whole) to preview what it would restore, and confirm it yourself; no order ever rolls back for you. A hooks-only session has no launch checkpoint, so there is nothing to roll back to.

agentgarm rollback <task>

See also

Undo what an agent changed

Preview what rolling back an agent's task would restore and keep, then apply the rollback from the checkpoint taken before the agent started.

5. Ask who can lift it

A kill order is lifted by its lift, which needs as many approvers as the order did: ask your Central Lite admins. If they and Central Lite cannot be reached, root on the machine can lift it locally with sudo agentgarm kill-order lift and a reason; that is recorded and reported to Central Lite. Root removes any other hold with sudo agentgarm task resume, naming the hold: breach, or the command: hold Central Lite placed (which its admin can also remove). Both run outside every governed task. A no-root install takes no kill order; there the daemon's own user runs task resume, at assurance C.

task resume answers that the hold is removed and the task is running again, or that it is still suspended by the holds it names; it refuses an order's hold. Resume and the local lift are not built on Windows yet: a suspended task there stays suspended until it is killed.

sudo agentgarm task resume <task> --hold breach
sudo agentgarm kill-order lift <order> --reason "approvers unreachable: lifted on the machine"

See also

Stop agents across a fleet

Suspend, terminate or lock down agents on one machine, one agent, one person or the whole fleet, with an order two approver keys sign and each node checks.