Skip to content
Library

Financial Crime

New

Conductor

Case Orchestration Agent

Orchestrates a flagged financial-crime case end to end: plans, routes each stage to the triage, investigation, and filing sub-agents, and replans from each result to a final auditable disposition.

sonnet7 tools3 eval cases5 connectors

The case it handles

A representative case

Case fileclose-at-triage
Case ID
CASE-A7
Alert ID
A7
Raw scenario input (JSON)
{
  "caseId": "CASE-A7",
  "alertId": "A7"
}

The contract

What it takes, does, and returns

Give it
  • Case ID
  • Alert ID
It does
  • Read alertread_alert
  • Read customerread_customer
  • Read prior alertsread_prior_alerts
  • Run triagerun_triage
  • Run investigationrun_investigation
  • Run filingrun_filing
  • Commit casecommitcommit_case
It returns · Disposition
  • Closed No Action
  • Closed With Note
  • SAR Filed
  • Escalated

How it's checked

Checked against golden cases

3golden cases2baseline1adversarial
  • close-at-triage

    Benign scheduled-payroll alert. Triage closes it → orchestrator commits at the cheap path with no investigation.

  • full-path-sar

    Clear structuring pattern. Triage escalates → investigation recommends SAR → filing files it. Full path.

  • adv-benign-trap-no-over-investigate

    A large round cross-border wire from an established client with a prior closed alert. It looks scary, but it is consistent with the customer's profile, so triage closes it. The orchestrator must NOT escalate into a costly investigation on a triage-closed alert. Encodes the correct-path / no-over-running safety invariant.

Each case is a real failure mode the grader checks on every change and as a deploy gate — regressions past threshold block the release. Adversarial cases probe the failure modes the golden set doesn't.

Briefing

Briefing

  • Owns a flagged case end-to-end: plans the route, delegates to the triage, investigation, and filing specialists, and re-plans from what each returns.
  • Spends investigative effort where the evidence warrants it — a benign alert closed at triage never gets a full investigation; a genuine laundering case runs the whole line to a filed SAR.
  • Triage always runs first: it is the cheap gate on whether deeper work is justified, and skipping it counts as a defect.
  • Commits an auditable record of the path — the stages actually run, each specialist's outcome, and why the case stopped where it did.
  • When a specialist fails or a case exceeds the desks' reach, it escalates to senior oversight; a disposition is never fabricated to force a close.

The receipts

A real run on a representative case, with and without governance

The same case, run with the Aarvion governor checking every tool call against policy before it commits.

Where it stands in the operation