How Bayer Crop Science built an agentic decision engine on Amazon Bedrock AgentCore

What happened

Bayer Crop Science built its Agentic Decision Engine on Amazon Bedrock AgentCore to give users a natural language interface to its Connected Decision Engine analytics platform. The system runs 22 specialist agents and resolves about 80 percent of user questions without engineer involvement in its first phase. The case study explains five architecture decisions and production preparations with AWS Unified Operations, including durable state, authorization boundaries, tool scoping, failure monitoring, and prompt caching that cut input-token cost by 20–30 percent.

Read next

• More changes in today’s briefingOlmoEarth v1.3 releases global satellite embeddings and a browser-based mapping explorer

The story so far

Bayer Crop Science built an Agentic Decision Engine on Amazon Bedrock AgentCore to let users query its Connected Decision Engine analytics platform in natural language. According to an AWS Machine Learning Blog case study, the system runs 22 specialist agents and, in its first phase, resolves about 80 percent of user questions without engineer involvement. The case study describes how Bayer designed the system and prepared it for production with AWS Unified Operations, including preserving state, setting authorization boundaries, limiting which tools agents can use, and monitoring failures. It also reports that prompt caching reduced input-token costs by 20–30 percent.