How Bayer Crop Science built an agentic decision engine on Amazon Bedrock AgentCore with AWS Unified Operations

Bayer Crop Science developed the Connected Decision Engine (CDE) to run, monitor, and reuse machine learning (ML) workflows while maintaining repeatability and governance. As adoption increased, understanding results became a challenge due to context fragmentation across various systems and tools. To address this, Bayer created the Agentic Decision Engine (ADE), a multi-agent, natural language interface built on top of CDE.

By employing Amazon Bedrock AgentCore, Bayer was able to focus on agent behavior rather than managing its own agent hosting and tool access. The ADE uses four layers: user experience, agent orchestration, specialist agents, and tools and data sources. Users interact with ADE through a React web front end, which connects to a Microsoft Entra ID identity proxy before reaching the agents.

The agents are deployed and hosted in the framework-agnostic AgentCore runtime, allowing the team to use LangGraph for orchestration and their preferred models. The ADE architecture consists of a supervisor agent, which directs questions to specialist agents, and various specialized agents that handle different tasks such as conceptual question answering, workflow status retrieval, troubleshooting, and generating visualizations.

By implementing this system, Bayer reduced the time spent on manual log-hunting and engineering support, enabling breeders to obtain answers to their inquiries in under 10 minutes.