Guide 7 of 61 · Getting started
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.
See whether AgentGarm's services are up, what assurance level each running agent session has and why, what a bypass alert means and what to do about it, and let doctor check the rest.
1. Check the services
agentgarm status is read-only and always exits 0, so read what it says. Its first lines name agentgarmd (root or no-root mode, and its socket) and agentgarm-session with its proxy address; a proxy marked “NOT answering” means no agent can be launched, since agentgarm run refuses to start one without the proxy; agentgarm doctor then says to restart agentgarm-session. A session “NOT RECOGNISED by agentgarmd” is not counted as AgentGarm's: its budget holds fail open and its events are dropped. Under it, “console call(s) from inside a task” counts the console calls the session refused (the local UI, and an on-demand inventory sweep): to a process inside a governed task, to a caller whose process it could not identify, and to one agentgarmd could not place inside or outside a task, which includes your own agentgarm ui while agentgarmd was down or not answering. The probe agentgarm run makes at each launch to prove that refusal is not counted. A number above 0 most often means an agent tried to reach the console; if agentgarmd was unavailable at the time, it may be your own calls. The self-check line counts the gaps this host had at start, and the policy lines show each policy file in force, skipped or rejected. On an enrolled machine, a kill order in force comes first preview: LOCKDOWN, “this machine is locked down: no governed agent starts or acts”, or a kill order line, with the order, who approved it and why (what it means for you). agentgarm doctor does not report orders.
agentgarm statusChecked by the walk: this command ran as written on a no-root Linux install.
2. Read each session's assurance level
Start the agent with agentgarm run and run agentgarm status from another terminal: each live task has a task line with its level, and under it every class its policy uses (commands, files, network, model calls, approvals and the rest) with the layer that enforces it. A new task shows “C (assessing)” until agentgarmd sees its sandbox, within seconds. When agentgarmd sees that a task's processes have all exited, its line says “ENDED” (“ENDED while assessing” if its level never settled) until agentgarmd drops it at its next check, every 5 seconds; “ROOT EXITED” means processes the task started still run. “SUSPENDED” preview means the task is stopped, with one line for each hold that keeps it so, such as airbag or sensitive_frontmost (what lifts each), a budget's breach or a kill order's order: (every kind of hold). The coverage blocks that follow cover only the agents found installed, seen in a task or given a policy; --all lists every adapter.
A, enforced, needs an OS sentinel that makes launching through AgentGarm unavoidable; the sentinels are built but not yet claimed, so this release rates nothing A. B, controlled, is the sandbox of a launch through agentgarm run or agentgarm wrap, which you could have skipped. C, observed, is detection and accounting only; every agent's hooks are C today (coverage). A session is rated by the weakest class its policy uses, so command rules, which only hooks can judge, make most sessions C today while their files are B wherever the sandbox's containment check passes. How the levels are set.
See also
Govern an agent with no hooksRun an agent AgentGarm has no hook adapter for inside the same sandbox and proxy, and know which of your policy's rules that leaves unenforced.
3. Open the assurance panel
agentgarm ui prints the dashboard's address, with a launch code that works once; --open opens it in your browser. The Overview lists each live task with its assurance level (selecting a task says why it is not higher) and the newest alerts beside it; its Alerts screen lists them all: bypass, assurance lowered, DLP, policy-source conflict, checkpoints and budget. An unknown level is shown as C.
agentgarm uiagentgarm ui --openChecked by the walk: this command ran as written on a no-root Linux install.
4. Act on a bypass alert
agentgarmd checks every live task every 2 seconds. A bypass alert lowers its task to C for the rest of the task's life, and agentgarm status shows why on its lowered: line; fixing the cause does not raise it again, so fix it and start a new task.
- User off-switch: a model routing variable you exported, such as
ANTHROPIC_BASE_URLin a shell Claude Code started, or the agent's hooks switched off in its settings (Claude Code'sdisableAllHooks: true, Codex's[features] hooks = false). Unset the variable for the launch, or remove the setting. - QUIC or UDP to port 443, and direct connections to a model provider's address: traffic the proxy cannot see. Inside
agentgarm runthe sandbox stops both on macOS; on Linux it stops QUIC but limits TCP only by port, and only from Linux 6.7, so a direct connection is still possible there and network rules rate C. The alert matters most for an agent started outside it, so start the agent through the launcher. Some providers share CDN addresses with other sites, so check the destination before calling a direct connection a bypass.
env -u ANTHROPIC_BASE_URL agentgarm run -- claudeChecked by the walk: this command ran as written on a no-root Linux install.
5. Run doctor
agentgarm doctor checks agentgarmd, agentgarm-session and the proxy, that agentgarmd recognises the session, the sandbox backend (Seatbelt on macOS, Landlock among the active security modules on Linux), presence, git for checkpoints, the catalog and agent versions, the system files AgentGarm writes for Factory Droid, Gemini CLI and OpenCode preview, on a Mac the computer-use airbag when one is linked or a computer-use task expects one preview (it fails when the airbag is not live or lacks Accessibility, with the System Settings link), and that you are not running it as root. On a no-root install, which writes none of those files, their failures are shown as warnings. Each problem comes with what to do about it. It exits 1 when a check fails; a warning exits 0. The sandbox check is a precondition only: the canary still decides at every launch.
agentgarm doctoragentgarm doctor --jsonChecked by the walk: these commands ran as written on a no-root Linux install.