Bayer Crop Science described on the AWS Machine Learning blog how it built an Agentic Decision Engine (ADE) on top of its Connected Decision Engine, the platform that runs hundreds of ML model configurations for plant breeding across crops, regions and teams. ADE is a natural-language interface with a supervisor agent that classifies each request, routes it to one of 22 specialist agents and merges answers when a question spans domains. Specialists include a troubleshooting agent that reads CloudWatch telemetry and workflow status through the platform API, a domain expert answering from retrieval-augmented documentation, and an agent that generates charts. The supervisor runs on Amazon Bedrock AgentCore runtime with the team's own LangGraph orchestration, checkpoints persisted in DynamoDB, session and feedback records in Aurora PostgreSQL, response caching in ElastiCache, and backend APIs exposed as MCP tools through AgentCore Gateway so authorization stays outside agent code. Bayer reports that ADE resolves about 80% of user questions in its first phase without an engineer in the loop, and that diagnosing a failed phenotyping workflow now takes under 10 minutes instead of more than an hour.
Our takeThe design detail worth copying is the state handling: the runtime session is used for active execution, while the resumable graph state is written to a durable store, so a scale-out event or restart does not strand a multi-step diagnosis. The 80% figure is self-reported and unscoped, since it says nothing about the difficulty mix of the questions or how failures were counted. The routing problem it solves is real though: with 22 specialists, a misrouted question fails quietly rather than loudly.