Bayer Crop Science has built the Agentic Decision Engine (ADE), a multi-agent natural language interface to its analytics platform. According to an AWS case study, ADE runs on Amazon Bedrock AgentCore and complements the Connected Decision Engine (CDE), which remains the governed analytics execution layer.

CDE deploys, orchestrates, monitors and reuses machine learning workflows while preserving execution history and operational controls. Its analytics cover genomics, field trials, imagery, environmental conditions and model outputs. ADE retrieves context previously scattered across APIs, documentation, metadata, dashboards and logs.

Specialists connected to operational context

ADE runs 22 production specialist agents. Domain Expert uses retrieval-augmented generation; Operational retrieves CDE workflow status and results; Troubleshooting analyzes telemetry; and Visualization supplies charts or dashboards. A Models as Skill pattern turns model code and documentation into centrally registered skills, grounding answers in implementation details.

A React frontend uses an Entra ID identity proxy. A LangGraph supervisor classifies requests, selects specialists and combines their results. GitHub Actions builds images, pushes them to ECR and deploys to AgentCore Runtime. Aurora PostgreSQL stores sessions, feedback and analytics, while ElastiCache caches responses. S3 content uses Titan Text Embeddings V2 and OpenSearch Serverless for retrieval.

AgentCore preserves context within a session. LangGraph checkpoints in DynamoDB let ADE resume after scaling, restarts or session boundaries. Entra ID supplies user identity, IAM authorizes service calls, and Gateway outbound authorization controls CDE access. A Lambda interceptor validates requests and passes authorization attributes. The team defined token contracts at each boundary with its AWS specialist.

Tool access and model compatibility

Registering OpenAPI specifications with AgentCore Gateway exposes REST operations as MCP tools without a server for each API. ADE offers about 20 API tools, scoped by specialist to limit input tokens and the tool selection space. Gateway also supports semantic tool search for larger catalogs.

CDE manages more than 500 model configurations. For ADE's release, Bayer retained its existing model after a newer version with reasoning enabled proved incompatible with its implemented combination of tool use and structured output. Production controls include least privilege, agent-scoped read-only tools and output validation. Troubleshooting cannot modify workflows, and Domain Expert cannot call unrelated APIs.

Visualization became a specialist when its work required retrieval, chart selection, retries, error handling and dashboard decisions. It uses matplotlib and Plotly in AgentCore Code Interpreter, exchanges data through S3 and can return Amazon Quick or Power BI assets.

Reported results and production safeguards

The source reports diagnosis time falling from over an hour to under ten minutes. When exception handlers made agent errors appear successful in monitoring, the team added CloudWatch failure metrics and alarms for tool errors, throttling, model timeouts and structured output validation. Traces and invocation logs span the supervisor and specialists.

ADE estimates interaction costs using inference profiles, request metadata, model invocation logs and IAM principal attribution. Visualization and Operational consumed the most tokens. Telemetry exposed repeated input without prompt caching; caching the stable prompt prefix reduced input-token cost by 20–30 percent, according to the source.

Bayer launched ADE to a gated group, with more than 100 breeders, analysts, researchers, data scientists and engineers using it. The account reports that about 80 percent of first-phase questions were resolved without an engineer, while Gateway reduced per-API integration overhead by more than 99 percent. These are reported outcomes of this deployment.

A designated AWS Technical Account Manager and Domain Specialist Engineer supported design through production, joining three technical standups weekly. Critical Workload and Well-Architected reviews with the Generative AI Lens informed stronger alarms, immutable versioned images, ECR scanning on push and image lifecycle policies. Load tests exposed an empty-response networking issue fixed before launch; 35 blockers were resolved.

Operational reviews continued after launch. Critical alarms are onboarded to AWS Incident Detection and Response; the source says incident engineers engage within five minutes of an alarm. AWS Security Incident Response supports security events. Bayer already has agent-to-agent communication and is exploring hybrid cloud interoperability. Evaluating Bedrock Guardrails alongside custom checks remains planned. Bayer retains ownership of the architecture and decisions.