The Agentic Bank

Credit Memo Agent

⬡ Scrivener Drafts the credit memorandum for committee from the diligence and spread.
◆ Assistive Orchestrator

Turns the spread and the diligence pack into a decision-ready credit memo: business overview, financial analysis, risk rating, structure, covenants, the recommendation and the case against it. Cites every figure back to its source, self-critiques the draft against a credit-policy judge over A2A, then hands the case to the chief-credit agent, which owns the recommendation, with a second agent re-deriving the sign-off.

Memory

Working The diligence pack, the spread, and the memo draft in progress.
Episodic Prior memos for the borrower and similar credits in the portfolio.
Semantic Credit policy, risk-grading methodology, memo standards and structure.
Procedural Memo-section templates refined from chief-credit-agent feedback.
Store Memo template library + credit knowledge base

Orchestration

orchestrator-worker MCPA2A

Harness · Managed Agents … long-running drafting session; structured note-taking for the memo outline; a credit-policy judge in the loop; compaction across sections.

Tools

{ } Spreading platform API Diligence pack retrieval Retrieval { } Risk-grading model API Credit-policy QA judge A2A

Evals & guardrails

  • A credit-policy agent-as-judge scores the memo for completeness + balance before review.
  • Citation discipline … every figure ties to the spread or a diligence source.
  • The chief-credit agent decides and a second independent agent re-derives the four-eyes sign-off; the memo agent never approves.
  • Sampled comparison of draft risk grades against final committed grades for drift.

Offline reflection

Replays chief-credit-agent feedback and grade overrides (Reflexion-style) to sharpen which risks the memo leads with and how the recommendation is framed.

Frontier edge

  • Long-horizon autonomy: drafts, self-critiques against the credit-policy judge, and revises a full credit memo across a multi-hour checkpointed session, not a single pass.
  • World-model simulation: models how the credit performs under the chief-credit agent's likely stress scenarios and writes the case-against from the simulated downside, not boilerplate.
  • Continual learning (eval-gated): self-edits memo-section templates from chief-credit-agent overrides, gated against grade-drift checks before the change ships … never auto-approving the credit itself.

A sample run

Trigger Diligence pack + spread arrive for the $40m regional-manufacturer facility.
  1. 1Draft the business overview and financial analysis from the spread, citing every figure.
  2. 2Pull the model risk grade; write the risk narrative, structure and covenant package.
  3. 3State the recommendation and the case against it.
  4. 4Run the draft past the credit-policy judge; revise flagged gaps.
Output A reviewed, decision-ready credit memo handed to the chief-credit agent, which interrogates and owns the recommendation, with a second agent re-deriving the sign-off.

In numbers

~1 hour
Median memo draft
140+
Memos drafted / day

Handoffs

Across ⇢ Risk → Credit Risk (model governance + final grade validation)

More on the Credit Underwriting & Portfolio Monitoring desk