Agentic Process Transformation

Patterns, accelerators, and hands-on workshops for building reimagined business processes with multi-agent workflows using:

• Strands Agents
• Amazon Bedrock
• Amazon Bedrock AgentCore
• Model Context Protocol (MCP)

Get started here

---

What's Inside

This repo has two main sections:

Agent Orchestration Patterns

A reusable pattern library for building individual agents and orchestrating them together. Each pattern is a self-contained Jupyter notebook that runs in Amazon SageMaker AI Studio.

Agent Patterns — how to build individual agents:

Pattern	Description
RAG Agent	Retrieves external knowledge to ground responses in real data
Tool Agent (Function Calling)	Agent calls internal functions — decides which, constructs arguments, executes
Tool Agent (MCP)	Agent delegates to an MCP server with its own runtime
Multimodal Agent	Processes PDFs, images, and mixed-format documents
Memory-Augmented Agent	Short-term and long-term memory across sessions
Supervisor Agent	Decomposes tasks, delegates to specialists, synthesizes results

Workflow Patterns — how to orchestrate multiple agents:

Pattern	Description
Routing	Classifies incoming work and dispatches to the right specialist
Parallelization	Fan out independent tasks to multiple agents, then aggregate
Human-in-the-Loop	Agent pauses at decision points for human approval
Reflect and Refine	Generator + reviewer loop with feedback until quality passes

Insurance Claims Processing Workshop

A complete end-to-end example: a life insurance death benefit claims pipeline built with four specialized agents. You start with a simple reasoning agent and progressively add tools, MCP integration, multi-agent orchestration, persistent memory, and human-in-the-loop escalation.

Agents:

Agent	Model	Role
Authenticator	Amazon Nova 2 Lite	Validates beneficiary identity and coverage via MCP
Extractor	Amazon Nova 2 Lite	Extracts structured JSON from 7 document types
Policy Verification	Anthropic Claude Sonnet 4.6	Cross-document consistency checks against policy terms
Communicator	Amazon Nova 2 Lite	Drafts claim decision notifications

Notebook sequence:

Notebook	What You Build
00_end_to_end_demo/	Interactive Streamlit app showing the full pipeline — run this first
01_a_simple_agent/	A simple Authenticator Agent with Amazon Nova 2 Lite reasoning
02_tool_augmented_agents/	Stub tools → MCP integration, document extraction with JSON schema
03_multi_agent_orchestration/	Intake Orchestrator with GraphBuilder, end-to-end pipeline
04_agent_core_integration/	Amazon Bedrock AgentCore Runtime (session isolation) and Memory (cross-phase context)
05_human_in_the_loop_integration/	AWS Step Functions callback pattern for human adjudication

Project Structure

├── 00-agent-orchestration-patterns/
│   ├── agent-patterns/               # 6 individual agent patterns
│   ├── agent-workflow-patterns/      # 4 multi-agent orchestration patterns
│   └── common/                       # Shared MCP servers and sample data
│
├── 01-insurance-claims-processing/
│   ├── agents/                       # 4 agent modules (factory pattern)
│   ├── mcp_servers/socotra_mock/     # Mock policy administration system
│   ├── notebooks/                    # Progressive workshop notebooks (00–05)
│   └── sample_data/                  # 7 sample claim PDFs

Key Technologies

• Strands Agents — Agent framework with tool orchestration and graph-based workflows
• Amazon Bedrock — Managed LLM inference (Amazon Nova 2 Lite, Anthropic Claude Sonnet 4)
• Amazon Bedrock AgentCore — Agent runtime with session isolation and persistent memory
• Model Context Protocol (MCP) — Standardized tool integration over stdio JSON-RPC
• AWS Step Functions — Human-in-the-loop callback orchestration

Target Audience

• Solutions Architects designing agentic workflows for business process transformation
• Software Developers and ML Engineers building agent-based systems
• Technical Leads evaluating agentic patterns for process automation

---

⚡ Getting Started — Deploy the Workshop Infrastructure

Prerequisites

• An AWS account with Amazon Bedrock model access enabled for:
  • Amazon Nova 2 Lite (us.amazon.nova-2-lite-v1:0)
  • Amazon Nova Multimodal Embeddings (amazon.nova-2-multimodal-embeddings-v1:0)
  • Anthropic Claude Sonnet 4.6 (us.anthropic.claude-sonnet-4-6)
• AWS CLI installed and configured with credentials
• Python 3.10+

Step 1 — Deploy the CloudFormation stack

This creates the Amazon S3 bucket, Amazon DynamoDB tables, IAM roles, and an Amazon SageMaker execution role with all required permissions.

aws cloudformation deploy \
  --template-file cfn-workshop-setup.yaml \
  --stack-name agentic-workshop \
  --capabilities CAPABILITY_NAMED_IAM \
  --region us-east-1

Step 2 — Get the S3 bucket name from the stack outputs

aws cloudformation describe-stacks \
  --stack-name agentic-workshop \
  --region us-east-1 \
  --query "Stacks[0].Outputs[?OutputKey=='S3BucketName'].OutputValue" \
  --output text

Step 3 — Upload sample claim documents to S3

aws s3 cp 01-insurance-claims-processing/sample_data/ \
  s3://<YOUR_BUCKET_NAME>/claims-processing/claimant-data/ \
  --recursive

Replace <YOUR_BUCKET_NAME> with the output from Step 2.

Step 4 — Set the bucket name as an environment variable

export WORKSHOP_S3_BUCKET=<YOUR_BUCKET_NAME>

Step 5 — Install Python dependencies and start Jupyter

python3 -m venv venv
source venv/bin/activate

# For the orchestration patterns
pip install -r 00-agent-orchestration-patterns/requirements.txt

# For the insurance claims workshop
pip install -r 01-insurance-claims-processing/requirements.txt

jupyter lab

Cleanup

To delete all workshop resources when you're done:

bash cleanup.sh

This script empties the S3 bucket (including versioned objects), stops running Step Functions executions, deletes AgentCore memories, cleans up IAM roles, deletes the CloudFormation stack, and removes the retained S3 bucket.

Note: If you prefer manual cleanup, delete the stack with aws cloudformation delete-stack --stack-name agentic-workshop --region us-east-1, then manually delete the S3 bucket: aws s3 rb s3://<YOUR_BUCKET_NAME> --force

Important Notices

Synthetic data only. All sample documents (PDFs, JSON mock data) in this repository are entirely fictional and generated for educational purposes. No real personally identifiable information (PII), protected health information (PHI), or financial data is included.

Not for production use. This workshop demonstrates agentic patterns for learning purposes. The automated claim adjudication shown here requires human-in-the-loop review for any production deployment involving financial decisions.

License

This library is licensed under the MIT-0 License. See the LICENSE file.