Multi-Agent A2A Implementation for Collaborative Financial Analysis (c-daniele.github.io)

A developer implemented an end-to-end, A2A (Agent-to-Agent) compliant multi-agent system for collaborative financial analysis, re-creating the AlphaAgents concept from a recent BlackRock paper (no code released). The system demonstrates framework-agnostic interoperability: an Agent Registry and standardized Agent Cards (.well-known/agent-card) allow dynamic discovery, while a Group Chat coordinator orchestrates specialized agents (Fundamental, Sentiment, Valuation) in a structured round‑robin debate to reach consensus. Risk-tolerance is automatically inferred from user queries, passed as metadata in JSON-RPC messages, and preserved through the debate to personalize recommendations. Technically, the prototype is built in Python 3.12 with FastAPI, JSON‑RPC for transport, LangChain Agents as a replaceable GenAI layer, and GPT‑4o for LLM calls. Key protocol and process choices include dynamic agent selection at runtime, a debate protocol enforcing minimum participation (≥2 turns per agent) and bounded rounds (≤5), cross-agent validation to reduce hallucinations, and consensus synthesis into a final report with debate metadata. Caveats: several practical simplifications (no RAG for fundamentals, keyword lookups on Yahoo Finance) mean outputs are illustrative, not investment advice. The work shows how A2A and registry-driven discovery can break monolithic agent frameworks and enable modular, extensible multi-agent ML systems.