How Bayer Crop Science built a multi-agent decision engine on Amazon Bedrock AgentCore

Bayer Crop Science built the Agentic Decision Engine, a multi-agent natural language interface over its Connected Decision Engine for ML workflows, hosted on Amazon Bedrock AgentCore with help from AWS Unified Operations. In its first phase it answers about 80 percent of user questions without an engineer involved.

The problem

Bayer runs hundreds of governed ML model configurations over genomics, field trial and imagery data. As adoption grew, understanding results became the bottleneck, because explaining a failure or a result meant piecing together APIs, documentation, metadata stores, dashboards and logs. Users needed to know which system to inspect.

How they did it

1. Place a supervisor agent in front of 22 specialist agents, each scoped to one domain with its own tools and data sources.
2. Run LangGraph orchestration on the AgentCore runtime and store checkpoints in DynamoDB so workflows can resume.
3. Register OpenAPI specs with AgentCore Gateway so REST APIs become MCP tools, scoped per agent.
4. Handle identity with Microsoft Entra ID for users, IAM for service calls and Gateway outbound authorization for CDE services.
5. Add custom CloudWatch metrics and alarms for tool errors, throttling, timeouts and output validation failures.
6. Attribute cost with Bedrock application inference profiles, invocation logging and IAM principals, then run reviews and load tests before a gated launch.

Results

• As reported by AWS, ADE resolves about 80 percent of user questions without an engineer in the loop.
• As reported by AWS, the Troubleshooting Agent cut a diagnosis from over an hour to under 10 minutes.
• As reported by AWS, caching the stable prompt prefix cut input-token cost by 20–30 percent.
• As reported by AWS, using AgentCore Gateway instead of a server per API cut per-API integration overhead by more than 99 percent, and more than 100 users have adopted ADE.

As reported by the source (AWS Machine Learning Blog customer case study); AgentGid did not measure these figures.

The agent used here

LangGraphLangChain · Agent frameworks and SDKs

A low-level orchestration framework for building stateful agents as graphs, with durable execution, persistence and human-in-the-loop steps. Availabl…

Price

Free (OSS)

Gid Score

72 #2 in Frameworks

What it does, pricing, alternatives →

Similar use cases

Recipe

Run LangChain's Open Deep Research multi-agent pipeline locally

LangChain's open-source deep research agent, built on LangGraph, runs research sub-agents in parallel, compresses what they find and writes a final report. It works with many mode…

On Deep Research Bench, the default GPT-4.1 configuration scored 0.4309 and the GPT-5 configuration 0.4943.

LangGraphResearch · Reporting

ShowcaseWeave

Route between coding-agent models with the open-source Weave Router 2.0

The Weave team presented Weave Router 2.0 in a Show HN post. It is an open-source routing model that sits in front of coding agents and switches between LLMs per turn, and the aut…

As reported by the Weave author on Hacker News: on Terminal Bench 4.0 and SWE Atlas, the router had equivalent pass rates to GPT-6 Astra.

Claude CodeOpenAI CodexOpenCodeWeave RouterCodexDevOps and CI · Workflow automation

Case studyDropsolid

How Dropsolid syncs CRM and marketing data with n8n workflows

Dropsolid, a Belgian digital agency built on open-source tools, adopted n8n to connect its CRM, marketing automation and analytics. The workflows sync data between systems and fee…

As reported by n8n (vendor case study), the workflows connected the team's apps and synced their data, resolving the original integration challenges.

n8nWorkflow automation · Reporting