Bayer Crop Science builds AI agent system to diagnose farming workflows

Solving the data bottleneck

Plant breeding at Bayer Crop Science requires analyzing massive datasets that include genomics, field trials, and environmental conditions. The company runs hundreds of different machine learning model configurations across various crops and regions, requiring every run to be repeatable and governed.

To manage this, Bayer created the Connected Decision Engine (CDE). This system allows researchers and data scientists to run analytics while keeping a consistent history of execution and operational controls. However, as the system grew, users struggled to understand the results because the context was spread across many different APIs, logs, and dashboards.

To fix this, Bayer built the Agentic Decision Engine (ADE). This is a multi-agent, natural language interface that sits on top of the CDE. Instead of users hunting through different systems, they can ask a question and have the AI navigate the operational stack to find the answer.

How the agentic architecture works

The ADE is organized into a layered architecture that separates the user experience from the underlying data. Users interact with a web front end, which passes requests to a supervisor agent. This supervisor acts as a router, classifying the user's intent and sending the request to one of 22 specialist agents.

Each specialist agent is scoped to a specific domain, such as troubleshooting, visualization, or domain expertise. For example, if a breeder asks why a workflow failed, the supervisor routes the request to the Troubleshooting Agent. That agent then reads telemetry from CloudWatch and retrieves status updates from the CDE API to provide a specific diagnosis.

The system uses Amazon Bedrock AgentCore to host these agents. This choice allowed Bayer to focus on agent behavior rather than building their own hosting layer. They used LangGraph for orchestration because it provides explicit control flow and checkpointing, which allows the system to resume multi-step workflows even if a session restarts.

Technical design and security

Bayer made several specific design choices to ensure the system could scale in an enterprise environment. One key decision was to externalize the orchestration state using Amazon DynamoDB. This ensures that the system can reconstruct or resume a workflow after a scale-out event or a session boundary without relying on a single execution environment.

Security is handled through a multi-layered approach that separates identity from authorization. While the company uses Microsoft Entra ID for user identity, they use AWS Identity and Access Management (IAM) for service-to-service calls. This prevents a single token from being reused across different trust boundaries, which helps secure the transition between user identity and tool identity.

The team also chose to scope the tools available to each agent. Because every tool definition consumes input tokens and increases the model's decision space, the system does not give every agent access to the full catalog. Instead, agents only see the specific tools relevant to their domain, which improves accuracy and reduces costs.

Operationalizing the AI

Moving the system into production required a focus on observability and cost management. Because a single user request can trigger a chain of multiple model calls, the team had to distinguish between transport success and task success. They added custom metrics to CloudWatch to track specific failures like tool call errors, model timeouts, and structured output validation failures.

To manage costs, Bayer implemented a strategy to attribute spending to specific agents and features. They discovered that the Visualization and Operational agents consumed the most tokens due to their reliance on multiple tool calls. By using prompt caching for stable instructions and tool definitions, they were able to cut input-token costs by 20–30 percent.

The system also uses a 'Models as Skill' pattern. This converts model code and documentation into skill definitions stored in a central registry. This allows the agents to answer questions based on the specific implementation of a model rather than relying on generic knowledge about the model's architecture.

Next steps for the platform

The ADE is currently resolving about 80 percent of user questions without requiring a human engineer to step in. This allows breeders and researchers to get answers in under 10 minutes, a process that previously took over an hour of manual log-hunting and waiting for support.

Bayer plans to further harden these controls by evaluating Amazon Bedrock Guardrails. This would add an extra layer of safety on top of the custom checks already built into the CDE workflows. These guardrails will complement the existing read-only tool access and structured output validation.

The team also continues to refine the system's ability to handle complex tasks. For instance, they recently moved visualization from a simple tool to a dedicated specialist agent. This change allowed the agent to own its own data retrieval, chart selection logic, and the decision of whether to generate a new chart or return an existing dashboard.