Skip to content

Architecture

Multi-agent architecture for financial services

The four-box vendor diagram is a slide. Production systems have a channel, a planner, specialists with tool rights, and a control plane that can say no.

Five layers, not four boxes

Channel: the customer app, the employee suite, the merchant plugin, or an MCP server. This is where identity is proven.

Orchestrator: the planner that decomposes a job (“dispute this card transaction”, “assemble a credit memo”) and routes steps. If this layer also holds the money-moving tool, you have a confused deputy.

Specialists: credit, KYC, disputes, inventory, portfolio, legal. Narrow, testable, versioned.

Tools: cores, card processors, document stores, market data, wallets. Every tool call is an audit event.

Governance: limits, allow-lists, human gates, model registry, prompt/version pin, kill switch. This is not a footer. It is a runtime.

Patterns that survive a risk committee

Maker-checker: the agent proposes, a human or a second model with a different prompt/tool set disposes above a threshold.

Read-then-write: retrieval and drafting are cheap; posting to a core is expensive and gated.

Bounded autonomy: refunds under a cap, disputes of a known reason code, KYC field extraction — not “do banking.”

Typed traces: every hop has an actor, a tool, a payload hash, and a duration. If you cannot replay a case, you cannot defend it.

What to measure

Containment, handoff quality, time-to-resolution, override rate, loss given error, cost per resolved intent, and — for money-moving agents — unauthorized-action rate. Accuracy of a demo prompt is not an operating metric.