AWS re:Invent 2025 - Moody’s: Architecting a multi-agent system on AWS (IND3303)

Moody's: Architecting a Multi-Agent System on AWS

Moody's Journey to Revolutionize Financial Intelligence

• Moody's, a 100-year-old credit rating agency, recognized the need to revolutionize their approach to serving customers in the financial services industry.
• Customers were asking complex, chained questions that went beyond the capabilities of a simple chatbot or single-model system.
• Moody's identified key challenges:
  • Diverse customer base with varying needs (commercial banks, asset managers, insurance companies)
  • Requirement for high accuracy (99%+) due to the high-stakes nature of financial decisions
  • Need to seamlessly integrate Moody's proprietary data with customers' own data
  • Difficulty in extracting insights from unstructured financial documents like PDFs

Evolving from Chatbots to Multi-Agent Systems

• Moody's started with a basic chatbot application (Research Assistant) in 2023, which provided fast, reliable answers grounded in Moody's research.
• However, this approach quickly hit limitations when faced with more complex, chained questions requiring cross-domain expertise.
• In 2024, Moody's introduced PDF upload capabilities, allowing customers to integrate their own data, but this highlighted the challenge of unstructured data.
• By 2025, Moody's had developed a sophisticated multi-agent system, leveraging specialized workflows, task-specific agents, and AWS infrastructure to deliver tailored solutions.

Architectural Decisions and AWS Services

1. Serverless Architecture:

   • Moody's chose a serverless approach to handle the spiky nature of financial services workloads.
   • This allowed them to scale up and down dynamically based on demand.

2. Modular Tools and Specialized Agents:

   • Moody's built a library of over 80 modular tools, each performing a specific task.
   • These tools were exposed to a variety of specialized agents, which could autonomously choose and orchestrate the necessary tools to solve complex problems.

3. Orchestration and Workflow Management:

   • Moody's developed a custom orchestration system to manage the complex, multi-step workflows (some with over 400 steps).
   • This orchestrator handled parallelization, cost optimization, and error handling to ensure reliable and performant execution.

4. Flexible Model Selection:

   • Moody's adopted a "buy vs. build" approach, leveraging AWS services like Amazon Bedrock to access a variety of large language models and embedding models.
   • This allowed them to choose the most appropriate model for each specific task, rather than being limited to a single model.

5. Unstructured Data Processing:

   • Extracting insights from complex financial documents, such as PDFs, was a significant challenge.
   • Moody's experimented with various approaches, including custom parsing algorithms and multi-modal foundational models, before settling on a multi-modal pipeline using AWS services like Bedrock Data Automation.
   • This pipeline intelligently routed different content types (text, tables, charts) to specialized processors, significantly improving the accuracy and scalability of their unstructured data processing.

Intelligent Document Retrieval and Agent-Assisted Workflows

• To enable users to effectively retrieve relevant information from Moody's vast repository of financial documents, Moody's developed an "Agentic Retrieval" system.
• This system decomposes user queries, creates a search strategy, executes multiple targeted searches, and then synthesizes the results to provide comprehensive, cited responses.
• Moody's also exposed many of their internal tools and capabilities as "Smart APIs", allowing customers to build their own agent-assisted workflows and applications on top of Moody's expertise.

Key Takeaways and Future Directions

• Moody's journey highlights the importance of moving beyond simple chatbots and prompt engineering towards true context-aware, multi-agent architectures.
• The use of AWS services like Amazon Bedrock, Bedrock Data Automation, and the upcoming Agent Core helped Moody's build a scalable, flexible, and reliable multi-agent system.
• Unstructured data processing remains a significant challenge in the financial services industry, and Moody's approach of using a multi-modal pipeline demonstrates a viable solution.
• Moody's is now exploring ways to further empower customers by exposing their internal tools and capabilities as "Smart APIs", enabling the creation of custom agent-assisted workflows and applications.

Technical Details and Metrics

• Moody's multi-agent system consists of:
  • 80+ modular tools
  • 100+ specialized workflows
  • Multiple task-specific agents
  • Processing over 1 million tokens per day
• Key AWS services utilized:
  • Amazon Bedrock for large language models and embedding models
  • Bedrock Data Automation for unstructured data processing
  • Upcoming Agent Core primitives for building production-ready agent systems

Business Impact and Real-World Applications

• Moody's multi-agent system enables their customers (commercial banks, asset managers, insurance companies) to make high-stakes financial decisions with unprecedented speed, accuracy, and context.
• By seamlessly integrating Moody's proprietary data with customers' own data, the system provides tailored insights and analysis that were previously difficult to achieve.
• The ability to extract insights from unstructured financial documents like PDFs is a game-changer, unlocking a wealth of valuable information that was previously inaccessible.
• Moody's "Smart APIs" empower customers to build their own agent-assisted workflows and applications, further extending the reach and impact of Moody's financial intelligence.

Examples and Use Cases

• A commercial bank using Moody's multi-agent system to analyze loan origination decisions, incorporating credit ratings, sector analysis, economic comparisons, and regulatory compliance.
• An asset manager leveraging the system to perform portfolio analysis, incorporating climate risk models and other Moody's data.
• An insurance company utilizing the system to validate regulatory compliance, drawing on Moody's expertise in this domain.