- AWS says Bayer's multi-agent assistant resolves about 80 percent of user questions about its ML workflows without an engineer, and cut some diagnoses from over an hour to under 10 minutes.
- The design lessons are about limits: read-only tools per agent, separate identity at each handoff, and monitoring that separates a completed request from a completed task.
- Ask your own team whether its agents hold separate permissions, whether monitoring can tell a wrong answer from a failed call, and who sees cost per agent.
A common test of an AI assistant is how many questions it answers. A better test is what the assistant is allowed to do while answering. AWS has published an account of how Bayer Crop Science built an internal agent system. It is most useful for that second question.
Running workflows got easier than understanding them
AWS says Bayer runs hundreds of machine learning model configurations across crops, regions and teams. Each run must be repeatable and governed. Bayer built a system called the Connected Decision Engine to manage them.
Then a new problem appeared. Users could run workflows, but they struggled to learn why a result looked the way it did. The answer sat in many places: APIs, documents, dashboards and logs. A user had to know which one to open.
Bayer's fix is the Agentic Decision Engine. A supervisor agent reads each question and passes it to one of 22 specialist agents. Each specialist covers a single domain. In the first phase, AWS reports, roughly four in five user questions need no engineer.
The post gives one example. A breeder asks why a workflow failed yesterday. A troubleshooting agent reads the workflow's telemetry and status, then names the failure point. Before, AWS says, that answer meant an hour or more of log-hunting and waiting for help. Now it takes under 10 minutes.
These figures come from a vendor's own case study. Also, "resolved" does not mean "correct."
How the system works
The agents run on Amazon Bedrock AgentCore, an AWS service for hosting agents. Bayer uses a framework called LangGraph to control the order of steps. AgentCore Gateway turns Bayer's existing REST interfaces into tools. It does this through the Model Context Protocol, a standard way for agents to call tools.
The team saved workflow state in Amazon DynamoDB, a database. An agent's session is not a durable store. With the state saved elsewhere, a restart should not lose a half-finished task.
Each specialist also gets only its own tools. Every tool description takes up tokens, the pieces of text a model reads. A shorter list leaves more room for the user's question.
The real design is a set of limits
The AWS post is most interesting where it describes restraint. The troubleshooting agent can read operational data but cannot change workflows. The domain expert agent can search its knowledge index but cannot call unrelated APIs.
AWS names four controls running in production. They are least privilege, read-only tools for each agent, checks on the structure of answers, and answers tied to retrieved data.
Identity gets the same care. Bayer keeps Microsoft Entra ID for users and uses AWS IAM between services. Access to downstream systems has its own mechanism. The post treats each switch between user, agent and tool as a likely weak spot. For every switch, it says, define who issued the credential, who it is for and how long it lasts.
This shows what an agent really is. It is a new kind of worker that needs a separate badge for each door, not one master key.
A green dashboard can hide a failed answer
Bayer hit a quiet failure. In some cases an agent failed, yet the application still sent back a reply. Monitoring scored that reply as a success. The call worked. The task did not.
The team reproduced the problem and added custom metrics inside the failure paths. It also added alarms for tool errors, throttling, model timeouts and failed answer checks. The post's rule is to separate a call that went through from a job that got done.
Cost has to be traced to an agent
One request can touch several models and tools. A single Bedrock bill therefore hides which part costs what. Bayer tied usage to individual agents. The visualization and operational agents, which make many tool calls, used the most tokens.
Telemetry also showed one workload that resent a large block of input with no caching. AWS reports a cut of 20 to 30 percent in input-token cost after Bayer cached the fixed opening of its prompts.
Questions for your own team
Bayer engaged AWS Unified Operations, which supplied a technical account manager and a generative AI specialist engineer. AWS says the team tracked 35 blockers to resolution before launch. This is AWS describing its own service.
Whatever vendor you use, put these questions to the team building your agents:
Does each agent hold its own permissions, and are they read-only where possible? Can monitoring tell a wrong answer from a failed call? Can finance see cost per agent rather than one monthly total? Is there a named owner for each handoff between identities?
An agent that can explain everything should be able to change almost nothing. Bayer's account describes limits built into the design: IAM roles scoped to each agent, tools scoped per specialist, and read-only access. AWS adds that Bayer has not yet adopted a managed guardrail service and intends to test one on top of these controls.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: AWS.
