Guide 52 of 61 · Running a fleet with Central Lite
Roll out policy from git with canaries
Keep agents.d policies in git and let Central Lite roll each signed commit out to canaries, then the fleet, halting by itself when machines report trouble.
Built, not yet claimed preview: this guide describes a package that has not passed its release gate.
Keep your fleet's agents.d policies in a git repository, sign each commit as a committer, and let Central Lite publish it as a signed bundle that reaches your canary machines first, then a quarter of the fleet, then every machine, halting by itself when machines it reached report trouble. At the end you can advance, halt or roll back a rollout, and replay a recorded day against a change before it reaches anyone.
1. Put your policies in a git repository
Central Lite publishes the policies of one directory of one repository at one commit: each <name>.yaml directly in it, named in lowercase letters, digits and hyphens, at most 256 of them. Subdirectories and other files are ignored, and a symbolic link with a policy's name refuses the commit. On every machine the bundle reaches, those policies replace its agents.d files as a whole.
agentgarm policy init writes a default pack, in audit mode, into any directory and needs no daemon; the Central Lite tarball carries no pack, so write it wherever agentgarm is installed. For a first try this guide keeps the repository beside Central Lite, made by the account that runs it; the commit and its note are made under your git identity, here the repository's own.
git init -b main policies
git -C policies config user.name "Alice Example"
git -C policies config user.email alice@example.com
agentgarm policy init claude-code --dir policies/agents.d
git -C policies add agents.d
git -C policies commit -m "Claude Code policy"Checked by the walk: this command ran as written on a Central Lite host.
See also
Watch your agent for a week, then enforceStart a policy in audit mode, see what enforcing it would have blocked in your own history, fix the rules that get in the way, then switch it to enforce.
2. Create a committer key and list it in the org root
Central Lite publishes a commit only with a committer's AgentGarm attestation over the repository, the commit and every policy text; git's own GPG or SSH signatures play no part. With no committer listed, it publishes nothing. garm-central policy committer-key --dir D creates the committer's Ed25519 key, or shows it, and prints its public key and key_id. With the default --keystore auto the key is in the macOS keychain, and elsewhere in a 0600 file under D/keys; a directory inside Central Lite's state directory is refused.
garm-central root lite --committer signs the next org root version listing that key, and every enrolled machine then checks the attestation too, so a stolen online policy key alone cannot push a policy nobody committed. Keep the committer key on a committer's own machine in production: under Central Lite's 1-of-1 root, whoever controls Central Lite's host can sign a root listing a committer of their own.
garm-central policy committer-key --dir committer
garm-central root lite --committer "$(garm-central policy committer-key --dir committer | awk '/^public/ {print $2}')"Checked by the walk: this command ran as written on a Central Lite host.
See also
Stand up Central LiteSet up Central Lite on one Linux host: its org and keys, its HTTPS port, a first admin and a backup passkey, and its first encrypted backup.
3. Point Central Lite at the repository and sign the commit
garm-central policy source set --repo takes an https:// or ssh:// URL, user@host:path, a file:// URL or an absolute path to a bare repository or a working copy; --ref defaults to main and --path to agents.d. A URL with credentials, http:// and git:// are refused, and a private https repository is not supported: use ssh or a local mirror. Central Lite pulls with the system git, which its container image does not have.
garm-central policy attest reads your local repository, runs Central Lite's lint gate, signs for --source-repo, which must be the repository exactly as policy source show prints it, and prints an agc1. line. --write-note adds that line as the commit's note under refs/notes/agentgarm-policy, where Central Lite reads it; with a remote repository, push the note with git push <remote> refs/notes/agentgarm-policy.
garm-central policy source set --repo "$PWD/policies"
garm-central policy source show
garm-central policy attest --repo policies --source-repo "$PWD/policies" --dir committer --write-noteChecked by the walk: this command ran as written on a Central Lite host.
4. Choose the canaries and the soak
Each publication starts a rollout in stages: the canary hosts, then enough hosts that at least a quarter of the fleet is reached, then every host, including those that enrol later. A stage advances once its soak has passed and at least 95 % of its members, the hosts that reported within the heartbeat window when it began, acknowledged the bundle and stayed healthy, or at its deadline if nothing went wrong. Hosts outside the stages reached keep the bundle in force before it.
garm-central rollout config sets what each new rollout copies: --canary takes a host id from node list (repeat it per host), and the defaults are a soak of 2 minutes, a deadline of 30 minutes and a heartbeat window of 2 minutes. With no canaries a rollout starts at the quarter. Without options the command prints the configuration.
garm-central node list
garm-central rollout config --canary <host id> --canary <host id> --soak 5mgarm-central rollout configChecked by the walk: this command ran as written on a Central Lite host.
5. Publish the commit and follow its rollout
garm-central policy publish pulls the repository now, stops on a policy that does not compile or any lint error, checks the attestation against the committers the org root lists, signs a bundle valid for 30 days and starts its rollout, printing the version and the rollout id. The same commit again is unchanged. A commit that does not descend from the newest one published is refused as a rollback unless you pass --allow-rollback --reason, which is chained. serve --policy-pull-every (at least 10s, off by default) publishes each new signed commit by itself and alerts policy_refused on a refusal. In the console, “Publish from the repository” under Policy repository in Policies and rollouts does the same, confirmed with your passkey.
A machine the rollout reaches checks the bundle's signatures, compiles every policy, and refuses the whole bundle if one fails, which counts against the rollout. Once in force, agentgarm status there shows each policy as central:<org> bundle <id> version <n>. rollout list shows the open rollout's state, stage and how many members acknowledged and are healthy.
garm-central policy publish
garm-central policy list
garm-central rollout listagentgarm statusChecked by the walk: these commands ran as written on a Central Lite host and a Linux node enrolled in Central Lite.
6. Advance, halt or roll back a rollout
Central Lite halts a rollout by itself, and alerts, when 5 % of the hosts it reached, and at least one, fail to load the bundle, restart, never acknowledge it, error on more than 1 % of decisions or answer too slowly, or when more than a fifth of a stage's members go silent. Denies are never a trigger. It rolls back by itself only when at least half the hosts reached, and at least two, fail to load or crash loop; otherwise a rollback is yours.
rollout advance moves to the next stage now, halt and resume stop and restart it, exclude --host sets one failing host aside from the health checks for good (it still gets the bundle; resume the rollout if that host halted it), and rollback re-issues the earlier content as a new, higher version to every machine, so no machine's anti-rollback mark moves down. In the console's Policies and rollouts section a running rollout offers Advance, Halt and Roll back, a halted one Resume and Roll back, and a completed one Roll back while it is the newest, each confirmed with your passkey; exclusions and the configuration stay on the host.
garm-central rollout show <rollout id>
garm-central rollout advance <rollout id>
garm-central rollout halt <rollout id> --reason "canary crash loop"
garm-central rollout resume <rollout id>
garm-central rollout exclude <rollout id> --host <host id> --reason "disk full"
garm-central rollout rollback <rollout id> --reason "breaks the build agents"7. Replay a recorded day before the change reaches anyone
Central Lite holds no decision inputs, so a replay runs over a day a machine recorded. agentgarmd replay export writes one user's decisions for one policy as a new 0600 file, holding only each action, its directory and what was taken; it is refused under the non-monitoring profile. Carry the file to Central Lite's host: garm-central policy replay --commit replays the configured repository's commit before publication against the newest bundle, with --home that user's home, and counts the decisions made stricter, looser or decided by another rule. The day is read, never stored.
With rollout config --replay-gate on (off by default) each new rollout waits, reaching no machine, until rollout replay releases it: at once when nothing is made stricter and every changed policy was replayed, otherwise only with --accept --reason, chained. In the console, “Replay and release” on the waiting rollout, confirmed with “Replay and release with a passkey”, does the same, for days small enough for one passkey step-up.
sudo agentgarmd replay export --policy claude-code --uid 1000 --day 2026-10-29 --out claude-code-day.jsonlgarm-central policy replay --commit --events claude-code=claude-code-day.jsonl --home /home/alicegarm-central rollout config --replay-gate on
garm-central rollout replay <rollout id> --events claude-code=claude-code-day.jsonl --home /home/aliceSee also
Test a policy before you ship itWrite what a policy must allow, ask about and deny as a file of events, replay it before you install the policy, and keep both in your repository.