Using the Console

Guard Policy

Policy is where you decide what your agents are allowed to do with money. The org ceiling applies to every agent; individual agents can narrow it further but never exceed it.

The fields

Spend cap — the total CSPR an agent may spend over the policy epoch. Enforced atomically, so concurrent requests cannot race past it.
Per-action max — the ceiling on any single authorization. Your defence against one catastrophic transaction.
Velocity / hour — how many authorizations are allowed per hour. Your defence against a fast loop.
Rails — which rails are permitted: casper-x402, cspr-trade, casper-deploy. Anything not listed is denied.
Service scope — the allowlist of services an agent may pay, e.g. svc:casper-paid-api. An unlisted counterparty is a DENY even if the amount is tiny.

Org ceiling vs. agent policy

Set the org policy on Guard Policy; set an agent’s own policy from its row on the agents screen. The effective policy for an agent is the intersection of the two — whichever is tighter on each field wins. Raising an agent policy above the org ceiling has no effect.

Choosing your first numbers

Set the per-action max at roughly the largest legitimate single call you expect.
Set the spend cap at what you would be willing to lose entirely.
Set velocity just above your expected steady-state rate — a runaway loop shows up as a velocity DENY long before it shows up as an empty float.
List only the rails and services the agent genuinely needs today.

NOTE

Policy changes take effect on the next authorization. Decisions already signed under the old policy remain valid and remain anchored — each decision records the policy reference it was judged against, so the history stays interpretable.

Kill switch

Suspending an agent stops it dead: every subsequent request is denied regardless of budget or policy. Use it first during an incident, then investigate in the Decision Stream.

Visual Builder

The Visual Builder lets you compose org ceilings, fleets, agents, guardrails, and delegated-key grants as a config graph. It is a design surface: edges express containment only, and deploying the graph to real, enforced policy is a separate, explicit action. Nothing you draw is live until you deploy it.

For the exact order in which these rules are evaluated at runtime, see The Policy Firewall.

← PreviousDelegated Keys Next →Decision Stream