Governing agents day-to-day

Once an agent is enrolled, its daily control surface is its detail page: open Agent Fleet at /dashboard/agents and click an agent — every control below lives there or one tab away.

Scope enforcement: Advisory vs Block

The Scope Enforcement card decides what happens when the agent presents a scope it isn’t authorized for. Advisory logs the violation but still verifies; Block refuses it. Either way the violation is audited, and enforcement only applies to scopes the operator declared.

1. Under This agent, pick Advisory, Block, or Inherit org default. The card shows the effective mode, marked (inherited) when no override is set.
2. Under Organization default, set the org-wide mode — it applies to every agent left on Inherit.

If the card shows Superseded by agent policy, a versioned agent policy governs this agent and both selects are disabled — manage it under Settings → Agent policy.

JIT credential activation windows

The JIT Credential Activation card (Scale plan) removes standing privileges: with the toggle on, the credential is suspended between task windows and the agent shows as Idle (JIT). Windows normally open from the operator’s own application via the wake endpoint.

• Default wake window — the TTL used when a wake call doesn’t specify one; blank falls through to your operator or org defaults.
• A running window appears as Active Task Window with its expiry and permitted scopes; end it early with Complete Task or Cancel.
• With no window open, Activate Now opens the Admin override: wake agent manually modal — break-glass only, recorded as an admin override in the audit trail; confirm with Wake (admin override).
• In Recent Task History, click Attest to record the post-task outcome — As intended, Partial, or Failed (the last two require discrepancy notes) — then File attestation; Amend revises it, keeping prior versions.

Fleet-wide, the JIT tab on /dashboard/agents shows the Operator wake endpoint snippet, the Recent wakes feed, and Organisation settings (the Max task TTL ceiling).

Model changes, drift, and re-verification

When an agent’s declared model is switched, or its runs report a different model family than declared, an amber banner appears on the detail page, below the Identity card — policies and role bindings matching the old model no longer apply to this agent. Click Re-verify now to open the credential reissue form (set the validity in days, then Reissue), or Update declaration if the change is intentional.

The org-wide toggle Require re-verification on model changes (bottom of the Scope Enforcement card): Enforced pauses verification and token minting until re-verified; Advisory gives the flag and webhook only.

Roles and scope-exception requests

Roles make agent authority reusable: assign them from the agent’s detail page, and manage the vocabulary on the Roles tab of /dashboard/agents — Organization roles (create one with + New role), My roles built from your own delegated entitlements, and the admin-only RBAC policy card. Out-of-role authority needs an approved exception: pending requests surface under Scope exceptions awaiting review — Approve, or Reject with a reason the requester sees. When a recertification cadence is set, due assignments appear under Recertification due; click Recertify and record a short reason. Roles can also auto-apply by attributes, tags, and operator relationships — see attribute & relationship access control.

The kill switch: suspend or retire

For a reversible stop, rely on suspension: a JIT agent with no open window is already inert, and an expired credential suspends the agent until you click Reissue Credential. To decommission permanently, click Retire Agent at the top of the detail page, optionally record a reason, and confirm with Confirm Retire. Retiring revokes the credential, fires the agent.retired webhook, and cannot be undone.