Bayer's 22-agent assistant answers 80% of questions without an engineer

Bayer Crop Science and AWS staff published an account on 23 October 2026 of an internal multi-agent assistant, the Agentic Decision Engine (ADE), running on Amazon Bedrock AgentCore. The authors say ADE resolves about 80% of user questions in its first phase without an engineer in the loop. More than 100 users across Bayer Crop Science have used it since launch.

What it does

ADE sits on top of Bayer's Connected Decision Engine (CDE), a platform for deploying and monitoring machine learning workflows used in plant breeding. The authors say the bottleneck had moved from running workflows to understanding them: why one failed, what a model did, where a result came from. A supervisor agent classifies each question and routes it to one of 22 specialist agents, each scoped to one domain with its own tools and data.

The post gives one example. A diagnosis of a failed phenotyping workflow that once took over an hour of log-hunting and waiting for engineer support now takes under 10 minutes, according to the authors. The stack uses LangGraph for orchestration, DynamoDB for checkpoints, Microsoft Entra ID for user identity, and about 20 API tools exposed through AgentCore Gateway.

What broke and what it cost

The most useful details are the problems. Application exception handlers caught agent errors and returned responses that monitoring counted as successes, so Bayer added custom CloudWatch metrics and alarms for tool errors, throttling, model timeouts and structured-output validation failures. Load testing exposed a networking failure that returned empty responses, which was fixed before launch. The team tracked 35 blockers to resolution.

On model choice, the team tested a newer model version with reasoning mode on, found its tool-use and structured-output combination incompatible with how the agents were built, and kept the existing model. On cost, one workload had a large repeated input prefix and no prompt caching; caching the stable prefix cut input-token cost by 20-30%, the authors report. The Visualization and Operational agents, which make the most tool calls, used the most tokens.

Controls

The authors say each agent's IAM role is scoped to its job: the Troubleshooting Agent can read operational data but not modify workflows, and the Domain Expert Agent cannot call unrelated CDE APIs. Responses are grounded in retrieved data and validated against a structured output. Bayer plans to evaluate Amazon Bedrock Guardrails but does not yet use them. Launch was to a gated group, not the whole organisation.

What the post leaves out

It is vendor-published, and it promotes AWS Unified Operations support. It gives no absolute spend, no cost per question, no accuracy measure beyond the 80% resolution figure, and no definition of how "resolved" was judged. The claim that using a gateway cut per-API integration overhead by more than 99% has no stated baseline.

What to do

• Alarm on task failure, not only on errors: test whether your monitoring would notice an agent returning a polite wrong answer.
• Give each agent only the tools it needs; Bayer scopes tool catalogues per specialist because every tool definition costs input tokens.
• Check for repeated prompt prefixes without caching before you negotiate model prices.
• Before swapping in a newer model, test tool calling and structured output together, not just answer quality.