AWS re:Invent 2025 - How Yahoo! Finance built research multi-agent systems with Gen AI (SPS321)

Introduction to Agentic AI and Multi-Agent Systems [00:00]

• Key Insights:

  • Rahul Gosh from AWS GenAI Innovation Center and Deepen from Yahoo Finance introduce the session focusing on enterprise adoption of multi-agent systems.
  • The session will cover building deep research agents, the capabilities of AWS Bedrock Agent Core, and Yahoo Finance's real-world implementation.
  • Key aspects include orchestrating agents, accessing multimodal data, managing diverse data sources, and evaluating performance.

  "AI agents represent a tectonic shift in how we build, deploy and interact with technology."

Foundations of Agentic AI [04:00]

• Key Insights:

  • AI agents are defined as autonomous systems that reason, plan, and complete tasks through iterative thinking and refinement, equipped with tools, memory, and agentic loops.
  • Four key enablers make agentic AI practical: model reasoning capabilities, data and knowledge integration, scalable infrastructure and agentic protocols, and agent development tools.
  • The agentic AI maturity scale ranges from rule-based automation to goal-driven AI agents and potentially agents that set their own goals.

  "So we see excitement across the industries and functions as these enablers converge to unlock the practical AI agent implementations."

Agentic Reasoning Patterns [07:36]

• Key Insights:

  • The core agentic loop involves an LLM with memory, tools, and goals, enabling it to reason and act iteratively.
  • The dominant pattern is ReAct (Reasoning and Acting), which involves intertwined loops of thought, action, and observation.
  • Other strategies include Review (reasoning without observation, generating execution plans upfront) and Reflection (promoting self-reflection and iterative refinement).
  • For deep research applications like financial analysis, a combination of these strategies is often employed for efficient data gathering, quality assurance, and adaptive reasoning.

  "For example, in financial research you have stock prices, earning reports, market indicators. Simultaneously it can call it can execute those steps and in a way it significantly speeds up the data gathering and these process in these comprehensive research workflows."

Single vs. Multi-Agent Systems and Patterns [10:47]

• Key Insights:

  • Single agents are capable, but expanding their scope too broadly can lead to mistakes and prompt bloat with too many tools.
  • Multi-agent systems are beneficial for complex use cases, allowing for agent specialization and improved overall performance.
  • Common multi-agent patterns include the Supervisor Sub-Agent pattern (hierarchical control) and the Agent Swarm pattern (collective and collaborative).
  • Hybrid patterns can combine structured hierarchical flows with collaborative intelligence.

  "So when you do decide to go into a multi-agent what are the different patterns that you have?"

Requirements of Financial Analysis Assistance [14:19]

• Key Insights:

  • Deep research in financial analysis demands complex reasoning, multi-step processes, hypothesis generation, data gathering, validation, and refinement.
  • Multimodal data integration is crucial, involving text, structured financial data, and visual/audio content from various global sources with varying reliability.
  • A scale versus quality trade-off exists, requiring a complete audit trail of data sources and material information detection while maintaining research quality.

  "So the first and foremost deep research demands complex reasonings. Here the financial analysis requires multi-step reasoning and hypothesis generation, data gathering and combining all of them through validation and refinement cycles."

Financial Research Architecture Example [15:51]

• Key Insights:

  • An example architecture using a Supervisor Sub-Agent pattern for financial research is presented.
  • It includes specialized agents for structured databases, news, and SEC filings, connected to a knowledge base (vector store).
  • Agents can interact with internal and partner APIs to gather information.
  • Each agent operates independently, coordinated by the supervisor for comprehensive analysis.

  "So in here first you can see an example of a structured database agent that is connected to your SQL databases has the SQL exeutor as a tool."

Challenges in Productionizing AI Agents [18:49]

• Key Insights:

  • Moving AI agents from development to production at scale presents significant complexity.
  • Challenges include infrastructure management, coordination between agents, resource management, security, governance, performance bottlenecks, and dynamic resource allocation.
  • Key operational challenges for enterprise-scale deployment are performance, scalability, security, context preservation across sessions, and auditability/compliance.

  "complexity around infrastructure management and everything will slow down the innovation."

Amazon Bedrock Agent Core Capabilities [24:14]

• Key Insights:

  • Amazon Bedrock Agent Core is a modular solution addressing production challenges for AI agents, offering a toolchain.
  • It provides a faster time to value by removing infrastructure and operational overhead.
  • Customers can use agents built on any framework or model, avoiding technology stack lock-in.
  • Key value propositions include secure, scalable, and reliable deployment of agents at enterprise scale.
  • Agent Core offers a secure runtime, agent gateway for tools, identity management, memory persistence, and observability.

  "So these are the primary challenges we at AWS has observed ourselves and that's why that's where this modular solution came in."

Deep Dive into Bedrock Agent Core Features [28:04]

• Key Insights:

  • Agent Core provides a secure and scalable runtime supporting multimodality (text, images, audio, video) and autoscaling for varying workloads.
  • It supports both synchronous and asynchronous inference, session isolation for multi-tenency, and integrates with existing observability tools via OpenTelemetry standards.
  • The platform simplifies integration with a simple 10-line code decorator for runtime, identity, and observability.
  • For deep research, it offers an 8-hour runtime for long-running workflows and advanced memory management with a dual memory architecture (long-term and short-term).

  "So essentially the runtime supports I think as Rahul mentioned a the DP such a requires the specific need for multimodality to do complex reasoning."

Yahoo Finance's Financial Question Answering System [34:21]

• Key Insights:

  • Yahoo Finance's mission is to empower investors with confident financial decisions, serving 150 million users.
  • A gap exists between retail investors (limited time, scattered info) and Wall Street professionals (teams, premium data, AI models).
  • The sheer scale of financial information (4,700 SEC filings daily, 110,000 press releases) makes it challenging for retail investors.
  • AI can help optimize strategies, process information, and discover opportunities by connecting dots impossible to link manually.

  "And that introduces a gap between these two cohorts of investors."

Challenges in Financial Question Answering [40:14]

• Key Insights:

  • Financial data is heterogeneous, with structured (price histories, financials) and unstructured (filings, news) components from diverse sources with different access mechanisms.
  • Temporal complexity arises from varying fiscal year patterns and the timing of financial events (e.g., earnings announcements).
  • Mixed-type datasets, like SEC filings, contain both structured tables and unstructured commentary.
  • Information flow varies, especially during earning seasons, requiring systems that can intake and make this data available for Q&A.

  "The first the data in the financial world is incredibly heterogeneous."

Architecture and Implementation on AWS [42:13]

• Key Insights:

  • The system evolved from a synchronous API to an asynchronous execution model using Lambda functions and SQS queues to decouple the API from agents.
  • Input guards filter out irrelevant queries before they reach agents, and output guardrails prevent financial advice or predictions.
  • The equity research agent acts as a supervisor, connected to specialized agents for financial data, insider transactions (tool-calling), SEC filings, and news (RAG-style).
  • Knowledge bases are updated via event-driven workflows, and observability solutions track metrics like token counts and latency.

  "So with that let's start looking into how we built the multi-agent system."

Utilizing AWS Services and Agent Core [46:32]

• Key Insights:

  • The architecture is designed to be cloud-agnostic, framework-agnostic, and model-agnostic.
  • On AWS, Lambda functions handle asynchronous execution, SQS manages message queuing, and LLMs are abstracted into services.
  • LangChain is used for tool-calling capabilities, while Bedrock Agents with Bedrock knowledge bases are used for SEC filings and news.
  • AWS Bedrock Guardrails are employed for input/output control. Agent Core simplifies this by integrating runtime, identity, gateway, memory, and observability features into a unified framework.

  "What can we do with agent core? It is a single framework that allows you to do everything uh with that."

System Examples and Guardrails [49:36]

• Key Insights:

  • Examples demonstrate the system's ability to compare company capex, explain stock price drops, and present insider selling information, even deriving new metrics like transaction value.
  • Guardrails are used to prevent the system from answering questions that constitute financial advice or market predictions, such as "Should I buy Apple?" or "Where is the market headed?".
  • AWS Guardrails help detect prompt injections and configure policies to filter queries based on topics, word policies, or sensitive information like credit card numbers and SSNs.

  "Couple of examples should I buy Apple where is the market headed this year Agent would refuse answering those questions because we have intentionally avoided that using guardrails."

Enhancing RAG and Evaluation [51:43]

• Key Insights:

  • Enhancements for RAG and knowledge bases include hierarchical chunking for large documents, integrating metadata (dates, filing types) for better retrieval, and query rewriting for time-sensitive queries.
  • Knowledge base reranking prioritizes relevant chunks.
  • System prompt refinements and named entity recognition help LLMs identify query intent and select appropriate tools.
  • Temporal complexity is addressed by providing a tool to query a company's fiscal quarter given its ticker.

  "Adopt adopt hierarchical chunking when you're working with large documents especially sec filings and news."

Evaluation Strategies and Results [53:41]

• Key Insights:

  • Evaluating financial Q&A systems is challenging due to the need for representative datasets and domain expertise.
  • A hybrid approach using both human and automated LLM-based evaluations is employed.
  • Templates are used to generate large question datasets, and a consistent rubric is applied for both human and AI judges.
  • The system achieves latencies of 5-40 seconds per query and has been tested for concurrency of 100 questions, with linear scalability.
  • Costs per query range from 2 to 5 cents, with token usage between 10,000 to 50,000 tokens.

  "So answers are produced by AI. AI is evaluating it. How do you trust that system? So we take a hybrid approach."