From your first agent to a decision you can inspect.

1. Define identity and scope

In Agents, create an agent, choose its domain, and configure allowed actions and resource scope. Use the agent-specific credential for authorization calls; the customer account key is not an agent identity.

Domain: insurance
Allowed action: approve_claim
Resource scope: claim-*

2. Ask before acting

Call the domain-neutral endpoint with an action, resource, and optional context. Action names must match your configured permissions exactly.

POST /api/v2/authorize?environment=sandbox
x-api-key: YOUR_AGENT_KEY
Content-Type: application/json

{
  "action": "approve_claim",
  "resource": { "id": "claim-123", "type": "insurance_claim" },
  "context": { "reviewType": "standard" }
}

3. Handle all three outcomes

allow: your application may execute the action. deny: stop and inspect the reason. escalate: hold the action while human review is pending. An HTTP 200 alone is not permission to execute.

Unconfigured and non-finance domains default to human review after scope checks. Finance has its own configurable autonomy policy; numeric context does not automatically enable autonomy in another domain.

4. Complete human review

Open Authorization Requests and select the request. Each active, eligible operator records a vote with a typed attestation. The request remains pending until quorum is met; rejection stops it immediately. Operators & Approvals summarizes coverage and pending work.

5. Inspect the evidence

Open Audit log for the request, votes, and final outcome. Drill into an audit event for read-only evidence, or use the audit explorer to verify the chain. Approval authorizes an action; your application still owns execution.