Bayer says its agents diagnose failed machine-learning workflows in the Connected Decision Engine (CDE). An October 23 AWS account reports gated use by over 100 staff.
The big change
- What changed: Agents answer questions about CDE records; CDE runs the analytics.
- Why it matters: Bayer reports about 80% of first-phase questions resolved without engineer help; no sample or independent accuracy check is provided.
- What to watch: For research-operations teams, the practical decision is which records an agent may read and how to tell whether its answer resolved the user's question.
The diagnostic route
Bayer describes a Microsoft Entra ID proxy, a LangGraph supervisor and 22 specialists. A troubleshooting specialist reads CloudWatch telemetry and CDE status through AgentCore Gateway with read-only tools.
AWS's OpenAPI target documentation explains how Gateway turns a described REST operation into an agent-callable tool. Its authorization guide separates access to the gateway from authorization at a downstream service. Bayer's permissions and diagnostic accuracy remain unverified.
Assessing the reported result
Bayer reports diagnosis falling from over an hour to under 10 minutes, without a measurement method. No crop or model outcome is measured.
AgentCore's observability documentation distinguishes built-in operational metrics from additional, application-specific telemetry. A research team evaluating a similar interface needs a task outcome measure: a successful software call by itself cannot establish that the user received the right explanation.
