A Deterministic Multi-Agent System for FinTech Compliance
A multi-agent KYC and transaction-compliance system for a highly regulated financial environment — engineered as a deterministic "software cage" around LLMs, so every decision is traceable to hard facts and legally defensible.
The challenge
Building a multi-agent compliance system is far more than chaining a few LLMs together. In a regulated banking and payments environment, the standard agentic playbook breaks down. Loading a 50-page Standard Operating Procedure into every prompt makes token costs skyrocket — but vaguely telling an agent to "follow industry best practices" makes it hallucinate compliance rules that don’t exist. And in compliance, a bare "Approved" or "Denied" from an LLM is legally insufficient: you have to explain exactly why.
Our approach
Deterministic tools, not prompt-stuffing
We shifted the heavy lifting — calculations, data aggregation, and retrievals — onto a Model Context Protocol (MCP) server and REST APIs, so the facts a decision rests on come from deterministic tools rather than the model’s memory.
Hyper-narrow sub-agents
Those deterministic tool outputs feed hyper-narrow sub-agents, each strictly restricted to analyzing a single perspective — no sprawling, do-everything agents that are impossible to reason about.
A “judge” agent with a clear audit trail
All sub-agent outputs funnel into one decision-making agent that must justify every verdict using only the hard facts extracted by the tools — producing a crystal-clear audit trail that explains exactly why a transaction passed or failed.
State & prompt management
We tamed authentication, authorization, and state across agents — a debugging bottleneck when passed through callback chains — and shifted the engineering discipline from versioning code to versioning prompts.
The outcome
The LLMs end up wrapped in a rigid, deterministic software cage: enough reasoning to analyze, none of the freedom to invent rules, and every decision traceable to hard facts. That makes the system fit for a highly regulated enterprise where legal defensibility — not model autonomy — is the point.
Technology
True enterprise AI in finance isn’t about granting an LLM autonomy — it’s about engineering a rigid, deterministic software cage around it to ensure strict legal compliance.