Velocity, daily budget, policy, colo bind, edge shield, and wallet lock sit in sequence on the pay path. No agent skips a layer. No spend bypasses the stack.
Every request is judged the moment it arrives at the Worker. Oversized bodies are rejected instantly. Global per-IP POST budgets stop floods before they reach any business logic.
No request proceeds without a verified identity. Either a human session cookie or a machine Bearer key must be present and valid. Everything else is rejected at the door.
Every API key can be bound to a Cloudflare colo and IP fingerprint. If the same key is used from a different edge location or network, the request is rejected before it reaches any business logic.
New keys start read-only. Spending power is explicitly granted through verbs, amount caps, and destination allow-lists. Anything outside the declared scope is rejected instantly.
Every agent is a real instance with a clear state machine. Paused or retired agents are blocked from spending or accepting work — even if their keys are still valid.
Daily spend limits, velocity windows, and automatic wallet locks protect against runaway agents. When thresholds are crossed, the system freezes spending and can suspend the key.
Every payment is evaluated against explicit policy. Allow and deny lists, maximum amounts, and task-prefix rules are enforced on the hot path before any balance is touched.
Beyond individual agent limits, the entire account is protected by daily and monthly fleet budgets. Soft warnings and hard stops keep total spend under control across every agent.
High-value, new-counterparty, or sensitive task payments can be forced into a human approval queue. The agent cannot proceed until a human decides — or the request expires.
When abuse patterns or key compromise signals appear, Edge Shield can quarantine an agent or freeze activity at the edge. This is the last automated defense before settlement.
│
