The engine room
The building cards show what each agent is meant to be. This is what the built ones actually are. 11 of this bank's agents are real and runnable on the Claude Agent SDK, and here they are opened up: the methodology the model follows, the sandboxed in-process tools it can call, the schema-validated outcome it must produce, the real data and regulations it stands on, and the three seams (model, tool, data) a runtime governance layer would clamp onto.
Real OFAC + FinCEN data · gated mock commits · an oversight judge on every filing
Financial crime
5 agentsSanctions, transaction-monitoring, investigation, and SAR filing. Real OFAC data and real FinCEN rules, an oversight judge gating every commit.
Screens against the real OFAC SDN list and never clears a strong match under unresolved doubt.
L1 triage that clears profile-consistent noise, escalates genuine typologies, and defaults to escalate under doubt.
Orchestrates an investigation and screens material parties by delegating to Cerberus over a real agent-to-agent call.
Drafts the SAR grounded only in the packet and real 31 CFR 1020.320; a thin packet yields a short, honest narrative, never invented specifics.
Turns CFR text into atomic obligations, maps them to the control inventory by scope (not keyword), and surfaces the gaps.
Consumer lending
6 agentsAuto and card decisioning, with the underwriter held structurally blind to prohibited basis and a fair-lending runtime guard on the model/tool seam.
Verifies the application, pulls the bureau file, and surfaces every data-quality and FCRA-dispute flag before underwriting.
The underwriter, kept structurally blind to prohibited basis: its tools expose none, and a fair-lending runtime guard redacts any that appears at the tool seam.
The one agent allowed to see the monitoring data, and only for outcome testing: it catches a disparate-treatment pair the underwriter cannot.
Derives the specific principal reasons for a denial only from the decision's own factors, grounded in real Reg B 1002.9 and the FCRA score disclosure.
Renders the consumer notice grounded only in the decision and the bank's disclosure profile (no fabricated creditor or CRA), validated before it commits.
A second-line audit in the spirit of SR 11-7: it tests whether the decision is supported and consistent, and catches an inconsistent decline.
Two oversight layers sit on top of these agents and gate every commit: an agent-as-judge on the financial-crime side (it independently reviews a SAR or a sanctions block before it commits) and the deterministic Keystone gate on the lending side (it holds a commit on a fair-lending violation, an unsupported decision, a model-risk fail, an unjustified policy override, or two coupled second-line concerns). They are described on each agent's page rather than given their own.