Skip to main content
A deployment puts published policy versions on a machine, each with one of two effects:
  • Observe records what the policy would have done, and blocks nothing.
  • Enforce acts on the decision: a deny blocks the call and an instruct steers the agent.

Add a machine

A machine appears under Admin → enforcement once it is connected to Cloud. If the one you want is not there yet:
  1. Go to Administration → Keys and create a key with policies:pull, so the machine can receive deployments, and events:add, so its decisions reach Cloud.
  2. Connect the machine with that key — Connect a machine to Cloud walks through it.
  3. Confirm it shows up under Admin → enforcement.

Deploy in observe mode

  1. Go to Admin → enforcement, find the machine, and expand its row.
  2. Select edit, add the tested policy version, and choose observe.
  3. Apply the change, then wait for the machine’s next check-in and confirm its deployment and coverage state.
  4. Go to Observe → policy to inspect live decisions. The machine deployment editor with policy versions, enforce and observe effects, and the apply deployment action.
1

Choose the version and machines

Pick the published version and the machines it should run on.
2

Observe

Review matches, reasons, affected tools, and false positives while nothing is blocked.
3

Enforce

Switch the effect to enforce once the observed matches separate unsafe actions from valid ones, then confirm every intended machine pulled the change and is reporting decisions.

Target an agent on a machine

Published policy versions do not contain agent-routing rules. By default, an assignment applies to all agents on its machine. When deploying a version, you can instead target an integration (including profiles added later) or a particular profile reported by that machine. The choice belongs to the machine’s assignment: the same published version can target different agents on different machines.
  1. In Admin → enforcement, expand the machine and select edit.
  2. On the policy’s agents control, keep all agents, select all hermes profiles (for example), or search for a reported profile by its label and agt_ ID. You can select multiple alternatives; a call matches if it belongs to any one of them.
  3. Review the target listed for each assignment in the deployment plan, apply the change, then check the machine’s reported deployment after its next check-in. History and rollback retain the selected targets.
The machine must have recently reported support for agent targeting before you can set a new target. You can select an exact profile only if that machine reported it with a hook installed and it is not stale. If the inventory is temporarily unavailable, existing targets remain assigned; they are not silently broadened to all agents. Reinstall older shell hooks to provide the profile hint used by new installations. After upgrading the CLI, run failproofai config to update and restart its daemon too: an older daemon cannot evaluate new profile-scoped calls and the hook refuses the mismatch. If the running agent’s profile cannot be identified (for example, when several settings scopes are installed but the hook does not report which one ran), targeted assignments do not match; unscoped policies still apply. Look for agent_scope_unresolved under that machine’s policy errors. A machine reports only the agents it traces: one entry per profile whose config carries the FailproofAI hook (Hermes and OpenClaw get one per profile, named by its label). An agent CLI that is merely installed is not listed. Each machine keeps up to 256 profiles in its roster. A profile’s ID never changes while it is listed. When the roster is full and a new profile with a FailproofAI hook appears (its hook runs, or its config carries the hook), the daemon makes room by dropping one profile that meets all of these:
  • it has no hook installed;
  • no assignment names it;
  • FailproofAI Cloud would not accept it as a new target.
The daemon prefers a profile no longer found on the machine, then the one unused longest. It never drops a profile an exact target uses, so those targets stay valid. If no profile can be dropped, the new profile has no ID yet, and targeted assignments do not match it until a slot frees up.

Deploy Jev checks

A Cloud policy can carry Jev checks as well as, or instead of, JavaScript. Its kind is jev (Jev checks only) or both (JavaScript whose verdict its own checks may clear). It deploys, toggles and rolls back exactly like any other policy. Jev checks run on FailproofAI Cloud; nothing is installed on the machine. The machine receives only the JavaScript of a both policy. While a jev or both policy is deployed to it, FailproofAI Cloud turns Jev on for the machine, and each checked tool call of an agent the policy targets goes to FailproofAI Cloud, which asks the checks deployed to that agent and returns their verdict. With none deployed, Jev is off there. There is no Jev mode to set by hand.
To trial Jev checks without them deciding anything, deploy the policy with :observe: its checks are asked and their answers recorded as “would have”, and they neither block nor clear anything. Redeploy it as enforce once the policy page shows what they would have done; then Jev’s checks block calls as well as clearing the JavaScript verdicts they review. A both policy deployed with :observe observes both halves. A both policy’s JavaScript can be cleared only by its own checks’ answers from FailproofAI Cloud. An installed pack’s check of the same name never clears it. If FailproofAI Cloud cannot be reached, or sets no mode that asks, the policy stays hard: every deny of its JavaScript stands. Agent targeting works for Jev-only policies too. For a both policy, the assignment’s one target governs both its local JavaScript and its Cloud Jev checks; a call outside that target runs neither half. Keep the policy at observe while trialing scoped checks, just as you would for an unscoped Jev rollout. A machine reports what stops it from asking, and fp fleet show <machine-id> lists it under policy errors:
  • transcripts_disabled: the machine was connected with --no-transcripts. Such a machine never sends tool calls to Jev.
  • jev_unconfigured: the machine’s key does not carry Jev.
  • jev_rate_limited: some calls were refused for the organization’s Jev rate limit and decided by the regex policies alone.
  • jev_unavailable: FailproofAI Cloud’s Jev failed several calls in a row, so the machine skips it for a minute at a time.
One Jev request carries a limited amount of question text. FailproofAI Cloud’s checks come before the checks of packs installed on the machine. A pack check that does not fit is left out and reported as jev_budget. A pack check with the same name as one of your organization’s Cloud checks is left out too, and reported as jev_name_clash: the Cloud check is asked in its place.

Check coverage

Coverage answers whether a policy is running where the risk is.
  1. Go to Admin → enforcement and review the enforcing and observing totals.
  2. Search for a machine by ID or label, or filter for machines missing a policy.
  3. Expand a row to compare assigned policies, reported deployment, last check-in, and history.
  4. Refresh after the machine’s polling interval when an applied deployment is still pending.
The Enforcement fleet showing policy coverage, machine deployment state, and observe and enforce assignments. Look for machines that never pulled the latest deployment, enrolled machines that stopped reporting, a policy assigned to the wrong environment, and version drift after an interrupted update. Label machines by workload and environment — hostnames alone rarely survive autoscaling or replacement:
Enforcement management is an administrative Cloud workflow. Do not treat root-only enforcement routes as ordinary customer /v1 API endpoints.