Overview

This article describes the development of an experimental multi-agent system for analyzing economic data from Federal Reserve, BLS, BEA and other sources using Anthropic's Claude models. The system processes publications, extracts economic indicators, and generates structured analysis reports.

• Final Product v2: AI Generated US Macro Analysis 12/29/2025 (90 day lookback)
• Latest AI generated Macro Reports (v3.0)

System Architecture

The system leverages CrewAI’s multi-agent framework to create a economic analysis pipeline with specialized roles and coordinated workflows.

Multi-Agent Framework (CrewAI)

The application uses CrewAI to coordinate four specialized agents in a sequential processing pipeline, where each agent's output becomes input for the next stage:

1. Global Macro Events Collector - Extracts data from Federal Reserve APIs, BLS, BEA and other sources asynchronously.
2. Global Events Analyst - Identifies relationships between economic events
3. US Impact Analyst - Creates economic analysis from perspective of US
4. Macro Analysis Verifier - Validates outputs against research standards

Language Models

The system distributes workload across different Claude models:

• claude-3-5-haiku-20241022: Used for data extraction (3,000-4,000 token limit)
• claude-sonnet-4-5-20250929: Used for analysis and verification (8,000 token limit)

This distribution prevents API overload while optimizing cost and performance.

Data Processing Pipeline

Phase 1: URL/API Processing

• Configurable processing of releases from ‘n’ most recent days (default: n=30)
• FOMC URL prioritization for latest monetary policy decisions