Part 3: Building an AI-powered assistant for investment research with multi-agent collaboration in Amazon Bedrock and Amazon Bedrock Data Automation | Amazon Web Services TechTricks365

In the financial services industry, analysts need to switch between structured data (such as time-series pricing information), unstructured text (such as SEC filings and analyst reports), and audio/visual content (earnings calls and presentations). Each format requires different analytical approaches and specialized tools, creating workflow inefficiencies. Add on top of this the intense time pressure resulting from rapidly evolving industry conditions and narrow decision windows. Delayed analysis can mean missed opportunities or failure to identify emerging risks, with potentially significant financial consequences.

AI-assistants boost productivity by automating routine data collection and processing tasks, surfacing relevant insights, and allowing analysts to focus on higher-value activities. However, a single AI agent struggles with complex, multistep investment research workflows to effectively handle the full spectrum of multiple specialized tasks. This is where multi-agent collaboration provides an advancement. By creating specialized AI subagents that excel in specific tasks and functionalities and using them together under a coordinated framework using an AI supervisor agent, we can address the full complexity of investment research workflows. A supervisor agent can intelligently decompose complex queries, delegate specialized tasks to subagents, and synthesize their outputs into comprehensive answers—similar to how a research team functions in the real world. The benefits of this approach are substantial, stemming from distributed problem-solving and specialization, improved accuracy through specialized expertise, enhanced scalability by adding new agent capabilities without rebuilding the entire system, and greater transparency because the reasoning process of each specialized agent can be tracked and verified.

Amazon Bedrock Agents uses the reasoning of foundation models (FMs), APIs, and data to break down user requests, gather relevant information, and efficiently complete tasks. With the multi-agent collaboration capability of Amazon Bedrock, you can build, deploy, and manage multiple AI agents working together on complex multistep tasks that require specialized skills. Amazon Bedrock Data Automation (BDA) enables the generation of useful insights from unstructured multimodal content such as documents, images, audio, and video for your AI-powered applications, and it can be used as a parser when setting up a knowledge base for Retrieval Augmented Generation (RAG) workflows.

In this post, we walk through how to build a multi-agent investment research assistant using the multi-agent collaboration capability of Amazon Bedrock. Our solution demonstrates how a team of specialized AI agents can work together to analyze financial news, evaluate stock performance, optimize portfolio allocations, and deliver comprehensive investment insights—all orchestrated through a unified, natural language interface. We use BDA as a parser when setting up a knowledge base for RAG workflows with multi-modal data. Amazon Nova understanding models are used as the large language models (LLMs) associated with the supervisor agents and subagents.

Check the video that demonstrates the solution,

Results from these agents are demonstrative, not to be taken as financial advice.

Solution overview

The multi-agent investment research assistant is composed of a supervisor agent and three subagents (quantitative analysis agent, news agent, and smart summarizer agent) working together in a coordinated framework, as shown in the following diagram.

The supervisor agent uses the instructions you provide to understand the structure and role of each collaborator agent. The supervisor agent orchestrates the overall investment research process by breaking down user prompts, delegating subtasks to specialized subagents, and consolidating their outputs to generate the final response. Each subagent is configured to carry out specialized tasks as follows:

• Quantitative analysis agent
  • Functions – Powered by stock-query and portfolio-optimization action groups.
  • Responsibilities – Query and analyze historical stock data and build optimized portfolio allocations based on user inputs such as stock tickers or investment amount.
• News agent
  • Functions – Powered by the web_search action group and the financial_analysis_KB knowledge base.
  • Responsibilities – Search and retrieve relevant financial data such as earnings reports and filings from the knowledge base for context. If information isn’t present in the knowledge base, construct a web query. Handles queries related to analyzing company financials, performance drivers, and management commentary.
• Smart summarizer agent
  • Functions – Uses LLM capabilities with specific prompt instructions, no custom action group.
  • Responsibilities – Takes in output from other subagents, such as recent news and financial data, and synthesizes the information into structured investment insights. Summarizes long texts or documents such as earnings reports into concise insights.

Multi-agent collaboration works as follows:

1. User prompts the multi-agent investment research assistant with a high-level research query.
2. Supervisor agent breaks the query into subtasks and orchestrates and invokes the relevant subagents in parallel or sequentially as needed:
   • News agent to retrieve financial reports or data
   • Quantitative analysis agent for stock data analysis
   • Smart summarizer agent to summarize lengthy texts
3. Supervisor agent consolidates the subagent outputs along with the original query context.
4. Supervisor agent passes the consolidated information to the LLM to generate the final research insights.

This architecture uses the strengths of the different specialized agents to comprehensively address the various aspects of investment research, and the supervisor agent provides the intelligent orchestration.

Technical architecture

The following diagram is the technical architecture.

Some of the key components and the associated codes of the technical architecture are:

• Amazon Bedrock Data Automation (BDA) for processing data – The following invokes a BDA job asynchronously using the invoke_data_automation_async() method from the bda-runtime client in boto3:

  bda_response = bda_runtime_client.invoke_data_automation_async(
              inputConfiguration={
                  "s3Uri": s3_uri_input
              },
              outputConfiguration={
                  "s3Uri": s3_uri_output
              },
          )

• Defining and building subagents with tools – The following creates a subagent using Agent.create. It defines two Lambda tools—one for fetching historical stock data and another for running portfolio optimization—then registers them with the subagent along with its role, goal, instructions, and the underlying FM.

  quantitative_analysis_agent = Agent.create(
      name="quantitative_analysis_agent",
      role="Financial Data Collector",
      goal="Retrieve real-time and historic stock prices as well as optimizing a portfolio given tickers.",
      instructions="""Specialist in real-time financial data extraction and portfolio optimization.
                          ...""",
      tools=[
          # Stock Data Lookup Tool
          {
              "code": stock_data_lookup_arn,
              "definition": {
                  "name": "stock_data_lookup",
                  "description": "Gets the 1-month stock price history for a given stock ticker, formatted as JSON.",
                  "parameters": {
                      "ticker": {"description": "The ticker to retrieve price history for", "type": "string", "required": True}
                  },
              },
          },
      ],
      llm=LLM,
  )

• Defining and building the supervisor agent (multi-agent investment research assistant) – The following creates a supervisor agent using SupervisorAgent.create. It defines the collaborator agents (subagents), and registers them with the supervisor agent along with its role, goal, instructions, and the underlying FM.

  investment_research_assistant = SupervisorAgent.create(
      "investment_research_assistant",
      role="Investment Research Assistant",
      goal="A seasoned investment research expert responsible for orchestrating subagents ...",
      collaboration_type="SUPERVISOR",
      instructions=f"""You are a Investment Research Assistant, a financial research supervisor overseeing multiple specialized agents. Your goal is to coordinate and synthesize their outputs to create a structured stock investment analysis.
                  ...""",
      collaborator_agents=[
          {
              "agent": "news_agent",
              "instructions": f"Always check the knowledge base (ID: {kb_id}) first. Use this collaborator for finding news and analyzing specific documents."
          },
          {
              "agent": "quantitative_analysis_agent",
              "instructions": "Use this collaborator for retrieving stock price history and performing portfolio optimization."
          },
          {
              "agent": "smart_summarizer_agent",
              "instructions": "Use this collaborator for synthesizing stock trends, financial data, and generating structured investment insights."
          }
      ],
      collaborator_objects=[news_agent, quantitative_analysis_agent, smart_summarizer_agent],
      llm=LLM,
  )

• Invoking the supervisor agent (multi-agent investment research assistant) – The following uses the invoke method to pass in the defined request as a prompt and get response from the supervisor agent, using the multi-agent collaboration architecture.

  request = """What's AMZN stock price doing over the last week
                 and relate that to recent news"""
  print(f"Request:n{request}n")
  result = investment_research_assistant.invoke(
      request,
      enable_trace=True,
      trace_level="core",
  )
  print(f"Final answer:n{result}")

Prerequisites

To deploy this solution, attach the appropriate permissions for Amazon Bedrock, AWS Lambda, Amazon OpenSearch Serverless, Amazon Simple Storage Service (Amazon S3), and AWS Identity and Access Management (IAM) to your execution role. You also need to deploy two AWS CloudFormation stacks: web_search and stock_data.

For more information on the permission policy, CloudFormation stacks, and setting up the environment, refer to the README.md in the investment_research_agent repository on GitHub.

Dive deeper into the solution

To dive deeper into the solution and the code shown in this post, refer to the investment research assistant agent GitHub repo. The repo contains instructions for the end-to-end solution, including setting up the supervisor agent and subagents, associated action groups, unstructured data (earnings reports PDF files and earnings call audio files), and structured data (stocks time series).

In the appendix at the end of this post, we share different questions asked by a financial analyst to the supervisor agent, the subagent invoked, and the answers from the supervisor agent.

Cleanup

To clean up this solution, make sure to delete all agents, knowledge bases, and associated Amazon OpenSearch Serverless instances. Delete both the web_search and stock_data stacks. This will make sure that no charges are incurred after you’re finished testing this solution. Read the associated notebook in the GitHub repository for more instructions on cleaning up the agents.

Conclusion

The multi-agent investment research assistant built using Amazon Bedrock Agents with its multi-agent collaboration feature represents how financial professionals can use AI to enhance their analytical capabilities. By using specialized agents—each focusing on distinct aspects of financial analysis—orchestrated by a supervisor agent, we’ve created a system that can process diverse data types, apply appropriate analytical techniques, and synthesize comprehensive insights in response to complex investment queries. The result is an AI assistant that can handle multistep financial analysis workflows that would otherwise overwhelm a single-agent system.

The application of multi-agent collaboration extends broadly across the financial sector, from equity research and portfolio management to risk assessment and compliance monitoring. A complex financial workflow that requires multiple specialized skills and data sources can benefit from this architectural approach.

The implementation uses key AWS services, including Amazon Bedrock for foundation models, the multi-agent collaboration capability of Amazon Bedrock Agents, Amazon S3 for document storage, Amazon OpenSearch Serverless for vector search capabilities, and Lambda functions for specialized actions. This serverless architecture provides scalability, and the knowledge base integration provides context from financial documents to ground the AI’s responses in factual data.

For more information:

• Get started building your own multi-agent investment research assistant by accessing our complete solution code and implementation guide in the GitHub repo.
• Explore how to customize the solution for your specific investment research needs by modifying the subagents’ specialized tools and action groups and by bringing your data securely to Amazon Bedrock Knowledge Bases.
• You can also explore and run Amazon Bedrock multi-agent collaboration workshop with AWS specialists or on your own.

Read more about single agent solutions for investment research with multi-modal data:

Read the blog post on Unlocking complex problem-solving with multi-agent collaboration on Amazon Bedrock for more details on benefits, design, and evaluation of multi-agent collaboration for enterprise applications.

About the Authors

Sovik Kumar Nath is an AI/ML and Generative AI senior solution architect with AWS. He has extensive experience designing end-to-end machine learning and business analytics solutions in finance, operations, marketing, healthcare, supply chain management, and IoT. He has double masters degrees from the University of South Florida, University of Fribourg, Switzerland, and a bachelors degree from the Indian Institute of Technology, Kharagpur. Outside of work, Sovik enjoys traveling, taking ferry rides, and watching movies.

Lucas Banerji is an AI/ML and GenAI specialist Solutions Architect at AWS. He is passionate about building agentic AI systems and exploring the frontier of what’s possible with intelligent automation. Lucas holds a degree in Computer Science from the University of Virginia. Outside of work, he enjoys running, practicing Muay Thai, and traveling the world.

 Mohan Musti is a Principal Technical Account Manger based out of Dallas. Mohan helps customers architect and optimize applications on AWS. Mohan has Computer Science and Engineering from JNT University, India. In his spare time, he enjoys spending time with his family and camping.

Appendix

The following table contains example questions asked and responses from the AI-powered assistant for investment research based on structured and unstructured data.