AWS re:Invent 2025 - Using Tools and Agents in Generative AI applications (TNC320)

Summary of AWS re:Invent 2025 - Using Tools and Agents in Generative AI applications (TNC320)

Limitations of Large Language Models (LLMs)

• LLMs have vast knowledge but it is static and limited to their training data
• LLMs cannot perform real-world actions like downloading files or accessing private data
• LLMs tend to "hallucinate" and provide made-up information when they lack the necessary data

Combining LLMs with Tools and Agents

• LLMs can be enhanced by integrating them with external tools and agents
• The agent acts as an intermediary, using the LLM's capabilities along with external tools to provide comprehensive responses
• Key components of this approach:
  • Prompt: Provides the agent's instructions and constraints
  • Model: Supplies the reasoning and language processing capabilities
  • Tools: Allow the agent to take actions and access external data sources

Retrieval Augmented Generation (RAG)

• RAG allows the agent to query its own knowledge base before going to the LLM
• This ensures the agent can provide facts and information from its own data sources
• Vector databases like Amazon OpenSearch enable fast and scalable knowledge retrieval

Frameworks for Building Agentic AI Applications

• Strands Agents: Open-source SDK that abstracts away agent deployment and monitoring
• Crew AI: Python-based framework that supports individual and multi-agent architectures
• LangChain: Focused on straightforward linear workflows, with support for memory and prompts
• LangGraph: Enables complex, stateful workflows with human-in-the-loop capabilities

Integration Protocols

• Model Context Protocol (MCP): Standardized way for agents to connect to various tools and data sources
• Agent-to-Agent (A2A) Protocol: Enables direct collaboration and task delegation between different agents

Key Takeaways

• LLMs alone have limitations, but can be enhanced by integrating them with external tools and agents
• Frameworks like Strands Agents, Crew AI, and LangChain provide abstractions and components to build agentic AI applications
• Standardized protocols like MCP and A2A enable seamless integration and interoperability between agents and tools
• This approach allows for more sophisticated, real-world AI applications that can access diverse data sources and perform complex tasks

Business Impact and Use Cases

• Financial research portal that combines LLMs, real-time data, and document analysis to provide comprehensive investment insights
• Intelligent recipe assistant that can search the web, access databases, and provide tailored cooking recommendations
• Travel planning agent that can coordinate with multiple specialized agents to plan a complete vacation experience

Technical Details

• Use of Amazon Bedrock models and serverless services like AWS Lambda for agent deployment
• Leveraging open-source frameworks like Strands Agents, Crew AI, and LangChain
• Integration with external data sources and APIs using standardized protocols like MCP and A2A
• Utilization of vector databases, real-time data feeds, and other external tools to enhance the agent's capabilities