AWS re:Invent 2025 - Integrate any agent framework with Amazon Bedrock AgentCore (AIM396)

Integrating Open-Source Agent Frameworks with Amazon Bedrock AgentCore

Overview

• Presentation on how to integrate open-source agent frameworks like LangChain, Hugging Face Transformers, and Anthropic's Strand with Amazon Bedrock AgentCore.
• Speakers: Shree Subramanian (AWS Principal Data Scientist), Aris Merchant (AWS Senior Specialist SA), and Keith Teare (VP Engineering, Cohere Health)
• Key focus on operational excellence and "agent ops" - evolving DevOps practices to support production deployment and management of agent-based applications.

Integrating Open-Source Agent Frameworks

• Bedrock AgentCore is AWS's end-to-end platform for building and operating agents at scale.
• Supports integration with popular open-source agent frameworks like LangChain, Hugging Face Transformers, and Anthropic's Strand.
• Allows deploying agents built with these frameworks to a scalable, serverless runtime with no/minimal code changes.
• Provides additional services for agent identity, memory, policy enforcement, and observability.

Agent Ops: Evolving DevOps for Agent-Based Applications

• Transitioning from traditional DevOps to "agent ops" - operational practices tailored for agent-based applications.
• Key pillars:
  1. Infrastructure: Leveraging managed services like Bedrock AgentCore to minimize infrastructure management.
  2. Tooling: Integrating open-source observability platforms like Lenzfuse for tracing, monitoring, and evaluation of agents.
  3. Processes: Adopting a structured approach to agent experimentation, testing, and production deployment.

Observability and Evaluation

• Importance of observability and evaluation for agent-based applications.
• Use of open standards like OpenTelemetry to instrument agents and export traces.
• Integration with Lenzfuse, an open-source LLM engineering platform, for advanced observability, prompt management, and offline/online evaluation.

Cohere Health Use Case

• Cohere Health is using Bedrock AgentCore and Lenzfuse to build a "Review Resolve" product for healthcare prior authorization.
• Challenges:
  • Precise information extraction from complex, domain-specific language.
  • Temporal reasoning and context management for clinical data.
  • Scalable evaluation of agent performance and prompt effectiveness.
• Key capabilities:
  • Integrating Bedrock AgentCore for serverless agent deployment and management.
  • Using Lenzfuse for observability, prompt management, and offline/online evaluation.
  • Leveraging Bedrock AgentCore's API gateway and identity services for security and compliance.

Key Takeaways

• Bedrock AgentCore provides a managed platform for integrating open-source agent frameworks and deploying agents at scale.
• "Agent ops" evolves DevOps practices to support the unique requirements of agent-based applications, focusing on infrastructure, tooling, and processes.
• Observability and evaluation are critical for agent-based applications, with integration of platforms like Lenzfuse.
• Real-world examples like Cohere Health demonstrate the practical application of these techniques in complex, high-stakes domains.