Bayer Crop Science Deploys Agentic Decision Engine on AWS Bedrock AgentCore
· Engineer's Notes · Cem Koyluoglu
Bayer built an agentic decision engine on Amazon Bedrock AgentCore to automate workflow diagnostics, reducing resolution time from over an hour to under 10 minutes.
What happened
Bayer Crop Science created the Connected Decision Engine (CDE) to run repeatable, governed ML workflows across genomics, field trials, imagery, and environmental data. To simplify interpretation of workflow results, Bayer added the Agentic Decision Engine (ADE), a multi‑agent natural‑language interface that routes user queries to specialist agents. ADE resolves roughly 80 percent of questions without engineer involvement and operates 22 specialist agents in production. The system runs on Amazon Bedrock AgentCore, leveraging its managed runtime, IAM‑scoped tool access via AgentCore Gateway, and framework‑agnostic support for LangGraph orchestration. ADE’s architecture comprises a React front‑end, identity proxy (Microsoft Entra ID), container deployment via GitHub Actions to Amazon ECR, and runtime hosting on AgentCore. Data stores include DynamoDB for LangGraph checkpoints, Aurora PostgreSQL for session records, ElastiCache for response caching, S3 for embedded content, and OpenSearch Serverless for retrieval‑augmented generation. Observability is provided by CloudWatch, and tool access is mediated through the AgentCore Gateway.
Why it matters in production
The shift from manual log inspection to an automated agent reduces diagnostic time from over an hour to under 10 minutes, directly impacting breeder productivity. Isolating each user session in the AgentCore runtime improves reliability by preventing cross‑session interference, while persisting workflow state in DynamoDB enables graceful recovery after scale‑out events or restarts. Separating identity management (Microsoft Entra ID) from service‑to‑service authorization (IAM) reduces the risk of token misuse across trust boundaries. Tool catalog scoping per specialist agent limits token consumption and model decision space, lowering latency and inference cost. Observability enhancements that distinguish transport success from task success allow precise detection of agent‑level failures, such as tool call errors or model timeouts, which were previously masked as successful HTTP responses. Cost attribution mechanisms combine Bedrock inference profiling with request‑level telemetry, providing unit‑level insight into multi‑model, multi‑tool requests that standard account‑level billing cannot reveal.
Engineering takeaways
- Persist multi‑step workflow state in a durable store (e.g., DynamoDB) rather than relying on runtime session memory to ensure resilience during scaling or restarts.
- Implement distinct identity and authorization layers: use enterprise identity (Microsoft Entra ID) for users, IAM for service calls, and AgentCore Gateway for downstream tool access.
- Scope tool catalogs to the specialist agent that needs them; avoid exposing the full tool set to every agent to reduce token overhead and improve response latency.
- Add custom observability metrics that capture task‑level outcomes (tool errors, model timeouts, output validation) in addition to transport‑level success indicators.
- When a capability grows beyond a simple bounded operation—requiring state, retries, or decision logic—promote it from a tool to a dedicated agent to encapsulate complexity and improve maintainability.
Related work on this site
- Agentic & LLM Automation — Multi-step LLM pipelines that run unattended: model cascades with fallbacks, validation gates, scheduled automation and alerting when something breaks.
- RAG & Grounded LLM Systems — LLM features that answer from the right source instead of from memory: document-grounded assistants, transcript-grounded chat and quality gates around generated output.
- YouTube AI Summarizer — Open-source Chrome extension on the Chrome Web Store: AI summaries, key points, deep analysis, a two-host AI podcast (Gemini TTS) and transcript-grounded chat for any YouTube video, using Groq or Ollama Cloud with the user’s own key.
- Automated AI News Pipeline (this site) — The pipeline behind this site’s Tech News: scheduled GitHub Actions scrape sources, an LLM cascade on Groq rewrites and enhances articles, and layered quality gates (date integrity, language checks, instruction-leak detection, duplicate detection) decide what gets published, with Telegram alerts.
This note was drafted with AI assistance from the primary source credited on this page, and automatically checked against that source before publishing.