The Agentic Bank

Credit Diligence Agent

⬡ Sherpa Assembles the diligence pack on a prospective middle-market borrower.
◆ Assistive Orchestrator

Spawns parallel research sub-agents across financials, industry context, peer comparables, news and litigation, and the borrower's existing exposure across the bank, then merges them into the cited evidence pack the credit memo rests on. Reconciles aggregated exposure against the single-name limit and flags concentration for the chief-credit agent.

Memory

Working The deal thesis and the evidence board being assembled.
Episodic Prior deals and the relationship's history with the bank.
Semantic Industry knowledge, comparable-company sets, the bank's risk appetite.
Store Knowledge graph (entity/exposure) + document store

Orchestration

orchestrator-worker MCPA2A

Harness · Managed Agents … orchestrator spawning parallel research sub-agents (financials, industry, exposure, news); fresh context per sub-agent, compaction on merge.

Tools

{ } Spreading platform API Industry + market data MCP { } Exposure / single-name limit system API News + litigation search Retrieval

Evals & guardrails

  • Citation discipline … every figure and claim links to a source record.
  • Agent-as-judge review of pack completeness vs. the diligence checklist.
  • Exposure aggregation reconciled against the single-name limit system.

Offline reflection

Replays which diligence threads moved committee decisions to prioritise the research that matters and stop padding the pack.

Frontier edge

  • Long-horizon autonomy: chains a full diligence build (financials, industry, exposure, news) as one checkpointed task.
  • Causal industry reasoning: builds a counterfactual 'what would break this borrower' thesis (rate shock, key-customer loss) rather than reciting peer multiples.
  • World-model simulation of exposure: simulates how the new facility shifts group-level concentration before the memo agent ever sees the pack.

A sample run

Trigger Coverage agent requests diligence on a $40m facility for a regional manufacturer.
  1. 1Spawn sub-agents: financial trend, industry outlook, peer comps, group exposure, news.
  2. 2Aggregate the bank's existing exposure to the borrower group against limits.
  3. 3Merge findings into an evidence pack with a draft risk narrative.
Output A cited diligence pack handed to the credit-memo agent, flagging concentration against the single-name limit for the chief-credit agent's attention.

In numbers

one overnight run
Diligence build latency
60+
Deals prepped / day

Handoffs

Across ⇢ Coverage agent (deal source) → chief-credit agent (decision)

More on the Middle Market Banking desk