Meridian Claims Intelligence

Incoming claim

Claim MC-2209-4471: rear-end collision on I-95 … press Process claim to run it through the three Meridian agents — every model call routed, metered and remembered by the control plane.

MERIDIAN AGENT PIPELINE

1 · Intake Agent triagewaiting

2 · Adjuster Agent reason.rcawaiting

3 · Settlement Agent summarizewaiting

Settle claim

Recommended amount from the Adjuster Agent.

Asks the plane: may Agent::meridian-settlement perform settle on Claim::<id>? Amounts over $10,000 are sent as risk_level: high. Nothing real is paid — settlement is simulated.

◆ WHAT THE CONTROL PLANE DID

Each agent is checked in the plane's agent registry (must be active), then asks POST …/economy/route for its capability — the router picks the cheapest capable model. Context flows through shared-memory delta-reads (GET …/memory/{claim}?layers=), not full-history resends. Every call is metered to the meridian ledger (POST …/economy/usage). Settlement is a Cedar policy decision (POST …/authz/evaluate) with a human approval when required.

Audit: every call is on the plane's audit chain for tenant meridian — routing decisions, usage records, memory layers, approvals.

TOKEN ECONOMY — THIS RUN (tenant: meridian)

Actual cost (priced by the plane)

$0.0000

Baseline (naive full-history pipeline)

$0.0000

Saved by shared memory + routing

—

SHARED OPERATIONAL MEMORY · scope = claim id

No layers yet.

SHA-384 hash chain: not yet verified

Meridian is a sample third-party application. It contains no AgentOps code — it consumes the agentic control plane over HTTPS with a tenant API key.
Every call is on the plane's audit chain for tenant meridian. Nothing real is paid, ever.