Eleven parallel agents examining letters of credit

An open source multi-agent system that examines Letter of Credit document presentations against UCP 600 banking articles and produces structured discrepancy reports, replacing a manual bank examination with an auditable workflow. Eleven specialist sub-agents run in parallel.

Problem

A Letter of Credit presentation is examined by hand against the UCP 600 banking articles. The examiner reads every document, checks it against the articles that apply, and writes up the discrepancies. It is slow, it does not scale, and the reasoning behind a finding lives in one person’s head.

Approach

I built a multi-agent system that examines a presentation against the UCP 600 articles and returns a structured discrepancy report. Eleven specialist sub-agents run in parallel, each responsible for its own slice of the rules, and their findings are aggregated into one report. The output is structured, so a finding can be traced rather than trusted.

Architecture

A document presentation fans out to eleven parallel specialist sub-agents, built on LangGraph over Amazon Bedrock, whose findings are aggregated into a structured discrepancy report served through FastAPI.

The hard part

Eleven agents in parallel is only useful if their findings compose into a single defensible report, so the aggregation step carries as much weight as the examination. Pydantic v2 models pin the shape of every finding, which is what turns eleven independent opinions into an auditable workflow instead of eleven separate opinions.

Outcome

The manual examination becomes a scalable, auditable workflow that produces the same structured report every run. It is open source, which is why this page can name the project outright and point at the code. Development continues as a personal project.

Next

Cutting an enterprise agent from two minutes to fifteen seconds