Bayer Crop Science builds multi-agent analytics engine on AWS

Bayer Crop Science has deployed a 22-agent analytics interface on Amazon Bedrock AgentCore, Bayer and AWS said in a joint technical account published Oct. 23. The system puts a natural-language layer over the company’s governed machine-learning workflows.

The companies said the system, called the Agentic Decision Engine, or ADE, has been used by more than 100 people across Bayer Crop Science since a gated launch. The companies also said ADE resolves about 80% of first-phase user questions without an engineer, but did not disclose a launch date, usage frequency or the basis for that rate.

ADE sits above Bayer’s Connected Decision Engine, or CDE, rather than replacing it. CDE deploys, orchestrates, monitors and reuses machine-learning workflows while retaining their execution history and operational controls. ADE uses a supervisor to classify each request, route it to the relevant specialist and combine results when a question spans domains. The specialists cover domain knowledge, live operations, troubleshooting and visualization.

Its four-layer architecture includes a React interface behind a Microsoft Entra ID identity proxy; LangGraph orchestration on AgentCore Runtime; the specialist agents; and the tools and data sources they use. DynamoDB stores LangGraph checkpoints, Aurora PostgreSQL holds session records, feedback and analytics, and ElastiCache caches responses. Content in S3 is embedded with Amazon Titan Text Embeddings V2 and indexed in OpenSearch Serverless for retrieval-augmented generation. CloudWatch receives application and runtime telemetry.

Bayer externalized LangGraph checkpoints to DynamoDB because AgentCore session context is not a durable workflow store. That design lets a workflow resume after a restart, scale-out event or session boundary without relying on the same execution environment. The team also separated identity boundaries: Entra ID supplies user identity, IAM authorizes service-to-service calls, AgentCore Gateway controls outbound access to CDE services, and a Lambda request interceptor validates requests and passes authorization attributes. AWS describes AgentCore Identity as providing authentication, authorization and credential management for agents and tools.

To control tool access, the team registered OpenAPI specifications with AgentCore Gateway so REST operations could be presented as Model Context Protocol tools. Each specialist receives only the tools relevant to its job instead of the full catalog. AWS documentation says the gateway can enforce policy-based authorization, guardrails and request or response interceptors outside the agent runtime while supporting IAM, caller IAM, OAuth and token-passthrough authorization modes.

That scoped-tool design provides a point of comparison with ShipStation’s AgentCore IT-helpdesk deployment, which applies separate safeguards to ticket actions. Bayer and AWS said ADE exposes about 20 API tools and that using the gateway instead of a dedicated server for every API reduced per-API integration overhead by more than 99%; the account did not disclose the calculation method.

The team treated model selection as a compatibility decision. Bayer and AWS said the team tested a newer reasoning-enabled model but retained its current model for the release because the newer option was incompatible with ADE’s combination of tool use and structured output. The account does not name either model.

The team revised production monitoring after application exception handlers returned agent errors in responses that the system counted as successful requests. The team added custom CloudWatch metrics inside failure paths and alarms for tool errors, throttling, model timeouts and structured-output validation failures. AgentCore’s observability service provides built-in metrics and CloudWatch-backed traces, logs and custom telemetry, including data on latency, token use and error rates.

For cost attribution, ADE combines Bedrock application inference profiles, request metadata, model invocation logs and IAM principal data to estimate interaction-level costs and identify expensive agents, according to Bayer and AWS. The companies said caching a stable prompt prefix lowered ADE input-token costs by 20% to 30%, without publishing an independent measurement or observation period.

Bayer and AWS said production responses use least-privilege IAM roles, read-only tools scoped to each agent, structured-output validation and grounding in retrieved CDE data and telemetry. Bedrock Guardrails remain under evaluation rather than being a deployed ADE control. Before launch, AWS-led workload reviews identified gaps in failure visibility, container-image immutability, traceability and vulnerability scanning. The team responded with stronger alarms, immutable versioned images, ECR scan-on-push and lifecycle policies.

Load testing covered the application and service quotas and exposed an empty-response networking failure that was fixed before launch, according to the account. Bayer and AWS said the team resolved 35 launch blockers. They also said ADE’s critical CloudWatch alarms are connected to AWS Incident Detection and Response, whose incident engineers engage within five minutes after an alarm, while AWS Security Incident Response supports potential or active security events.