Manufacturing RCA Agent

An AI-powered root cause analysis agent for manufacturing quality investigations. Built on AWS Bedrock AgentCore with the Strands Agents framework.

Problem

Manufacturing quality investigations require processing large volumes of data scattered across many sources — deviation reports, equipment logs, batch records, supplier quality data. These investigations are manual, time-consuming, and often stall because of the fragmented nature of the data. This agent automates the investigation process by consolidating knowledge from multiple sources and applying a structured root cause analysis methodology, reducing investigation time from days to minutes.

Architecture

graph LR
    Client["Invoking Client<br/>(AgentCore CLI / AWS SDK)"]
    AgentCore["Bedrock AgentCore"]
    Agent["RCA Agent Container<br/>(Strands Agents)"]
    Bedrock["Amazon Bedrock<br/>Claude Opus 4.6"]
    HybridKB["Hybrid KB<br/>(OpenSearch Serverless)"]
    GraphKB["GraphRAG KB<br/>(Neptune)"]

    Client -->|"invoke"| AgentCore
    AgentCore --> Agent
    Agent -->|"InvokeModel"| Bedrock
    Agent -->|"Retrieve"| HybridKB
    Agent -->|"Retrieve"| GraphKB

Loading

The agent runs as a container on Bedrock AgentCore. On each invocation, it calls Claude Opus 4.6 via Bedrock for reasoning and retrieves manufacturing data from two Knowledge Bases — one using hybrid search (OpenSearch Serverless) and one using GraphRAG (Neptune). Both KBs are queried together: hybrid search excels at keyword-precise lookups (e.g. lot numbers, equipment IDs), while GraphRAG discovers cross-document relationships between entities (e.g. tracing a defect back through suppliers, equipment, and process changes). Infrastructure is deployed via CDK; the agent container is built and pushed to ECR automatically by the AgentCore CLI.

Investigation Pipeline

The agent follows a structured 3-step methodology modeled after manufacturing root cause analysis standards (e.g. 8D, Ishikawa). Each step executes as a separate invocation to keep within timeout limits and allow inspection between steps:

Incident → Problem Definition → Root Cause Analysis → Verification → Root Causes

1. Problem Definition — Identifies affected items, establishes timeline, classifies items as reference (known good) vs study (potentially affected)
2. Root Cause Analysis — Queries Knowledge Bases to identify potential causes with supporting evidence
3. Verification — Verifies each cause by comparing reference vs study items, determining a RETAINED or ELIMINATED verdict

Each step is driven by an SOP (Standard Operating Procedure) defined in app/RCAAgent/sops/. The SOPs define the methodology — what to query, how to classify, what output format to produce.

Sample Data

The data/ directory contains 40 synthetic manufacturing documents (txt, csv, pdf) that are automatically uploaded and indexed into the Knowledge Bases during deployment. They simulate a realistic manufacturing environment across 13 document categories — batch records, QC reports, deviation reports, equipment logs, supplier quality data, risk analyses, maintenance records, and more. This represents the types of documents a quality engineer would typically search through during a real investigation.

Prerequisites

• AWS account with Bedrock model access enabled (Claude Opus 4.6)
• AWS CLI configured with valid credentials
• CDK bootstrapped (cdk bootstrap aws://ACCOUNT_ID/REGION)
• Node.js with CDK CLI (npm install -g aws-cdk)
• Python 3.12+
• uv (Python package manager)

Note: The agentcore CLI is installed automatically by uv run deploy if not already present.

Note: This sample targets EU regions (default: eu-central-1). The model inference profile (eu.anthropic.claude-opus-4-6-v1) is EU-specific. If deploying in another region, update the MODEL_ID environment variable (e.g. us.anthropic.claude-opus-4-6-v1 for US regions).

Quick Start

# 1. Set your AWS account ID (find it with: aws sts get-caller-identity --query Account --output text)
vim agentcore/aws-targets.json

Replace the placeholder with your 12-digit account ID:

[
  {
    "name": "default",
    "account": "123456789012",
    "region": "eu-central-1"
  }
]

# 2. Deploy everything (infra + data + agent + smoke test)
uv run deploy

The CLI deploys CDK infrastructure (S3, OpenSearch Serverless, Neptune, Knowledge Bases), uploads sample manufacturing data, deploys the agent via AgentCore, and runs a smoke test.

Running individual steps

uv run deploy infra     # Deploy CDK infrastructure only
uv run deploy ingest    # Upload data + wait for ingestion
uv run deploy agent     # Deploy AgentCore agent + patch permissions
uv run deploy test      # Run smoke test
uv run deploy destroy   # Tear down everything

Testing the agent

Quick one-shot validation:

agentcore invoke '{"step": "step1", "prompt": "Investigate quality deviation on product NEXUS CK MB Reference 30421. Lot 1009900940 reported calibration error B18 on primary signal exceeding 550 SIU specification limit."}'

Interactive 3-step investigation:

python scripts/invoke_agent.py

======================================================================
  Manufacturing RCA Agent
======================================================================

  Describe a quality issue and I will investigate it using
  a 3-step root cause analysis methodology.

  Examples:
    "Defect found in production batch BATCH-2026-001"
    "Quality deviation on assembly line 3, reported 2026-03-10"
    "Customer complaint about product ABC-123"

  Type quit to exit.
======================================================================

You>

Local Development

cd app/RCAAgent
uv sync --dev
AWS_REGION=<your-region> KB_HYBRID_ID=<hybrid-kb-id> KB_GRAPHRAG_ID=<graphrag-kb-id> agentcore dev

The dev server starts at http://localhost:8080/invocations with hot reload.

View runtime logs:

agentcore logs                       # Stream logs in real-time
agentcore logs --since 1h            # Search historical logs

Configuration Reference

Variable	Description
MODEL_ID	Bedrock model ID (default: eu.anthropic.claude-opus-4-6-v1)
RETRIEVAL_MODE	hybrid, graphrag, or both (default: both)
LOG_LEVEL	Logging level (default: INFO)

Teardown

uv run deploy destroy

Security

See CONTRIBUTING for more information.

License

This sample code is made available under the MIT-0 license. See the LICENSE file.

Disclaimer: This is sample code for demonstration and educational purposes only, not for production use.