Agentic AI on AWS Workshop

한국어 | English

Build an AI agent from scratch with the Strands Agents SDK, then deploy and operate it with Amazon Bedrock AgentCore.

• How you learn: each chapter has a labs/ folder with empty files that you fill in yourself, and a completed/ folder holding the reference implementation. You write the code, then compare against the reference.

• Level: 100 to 200 (beginner to intermediate). No prior agent or LLM experience required.

• Duration: about 2 hours for the required chapters, about 3 hours for all eight lab chapters.

---

📚 What you will learn

• Strands Agents SDK: build an agent from a prompt, a model, and tools, then extend it with custom tools and MCP servers
• Retrieval: query an Amazon Bedrock Knowledge Base from an agent with the retrieve tool
• Multi-agent systems: Agents-as-Tools, Swarm, and Graph, and when to reach for each
• Memory: short-term and long-term memory with Amazon Bedrock AgentCore Memory
• Deployment: turn a local agent into a serverless one on AgentCore Runtime with four added lines
• Observability: agent metrics, logs, and OpenTelemetry traces, both self-managed and through CloudWatch GenAI Observability

---

🗂️ Chapters

#	Chapter	What you build	⏱️ Time	📊 Level	Track
00	Setup	Python environment, AWS credentials, Bedrock model access	10 min		Required
01	Single agent	Agent with prompt, model, and tools. Bedrock Knowledge Base, MCP tools, self-improving agent	30 min		Required
02	Multi-agent patterns	Agents-as-Tools, Swarm, and Graph	30 min		Required
03	Chatbot application	Streamlit chat UI with streaming and tool-call display	10 min		Optional
04	Observability with Strands	Metrics, logs, and OTLP traces to a local Jaeger	30 min		Optional
05	Agent memory	Short-term and long-term memory with AgentCore Memory	30 min		Required
06	AgentCore Runtime	Serverless deployment of the agent	20 min		Required
07	AgentCore Observability	CloudWatch GenAI Observability dashboard	10 min		Required
08	Developing with Kiro IDE	Steering, MCP config, and spec-driven development	10 min		Optional

Important

Chapters 01, 02, 05, 06, and 07 form the core path. Chapters 03, 04, and 08 are self-contained and can be skipped. Chapter 07 is the one dependency worth noting: it reads telemetry from the agent you deploy in chapter 06.

---

🚀 Quick start

Prerequisites

• An AWS account with permission to call Amazon Bedrock and create AgentCore resources
• Amazon Bedrock model access enabled in us-west-2 for the Anthropic Claude models listed below
• Python 3.12
• uv
• Docker, for chapters 04 and 06 only
• AWS CLI configured (aws configure), default region us-west-2

Install

git clone https://github.com/aws-samples/sample-aws-agentic-ai-workshop.git
cd sample-aws-agentic-ai-workshop/code/00-setup
uv sync
cd ..

Run your first agent

uv run --project 00-setup python 01-single-agent/completed/basic.py

If that prints an agent response, your environment is ready. Now open the setup guide for the full setup notes, then start chapter 01.

Note

All lab commands are relative to the code/ directory, so run them from there.

---

🤖 Models and region

The labs run against Amazon Bedrock in us-west-2. Enable model access for these before you start:

Model ID	Used in
us.anthropic.claude-sonnet-4-20250514-v1:0	Chapters 01 to 06
us.anthropic.claude-sonnet-4-6	Chapter 01 self-improving agent labs, chapter 02
us.amazon.nova-pro-v1:0	Chapter 04, metrics lab

Enable them in the Bedrock console under Model access. Cross-region inference profiles (the us. prefix) require access in the destination regions of the profile, which the console handles for you.

---

🛠️ Technologies and services

Technology	Purpose	Used in	Documentation
Strands Agents SDK	Agent framework	All chapters	Docs
Amazon Bedrock	Managed model inference	All chapters	Docs
Bedrock Knowledge Bases	Managed RAG over your documents	01	Docs
Model Context Protocol	Tool integration standard	01, 08	Docs
AWS MCP servers	Ready-made MCP servers for AWS	01, 08	Docs
Streamlit	Chat UI	03, 05	Docs
OpenTelemetry, ADOT, Jaeger	Trace collection and viewing	04	Docs
AgentCore Memory	Short-term and long-term agent memory	05	Docs
AgentCore Runtime	Serverless agent hosting	06	Docs
AgentCore Observability	CloudWatch GenAI Observability	07	Docs
uv	Python environment and dependencies	All chapters	Docs
Kiro	AI-powered IDE	08	Docs

---

📁 Repository layout

Code lives under code/, and the lab guides live under workshop/. The two trees mirror each other chapter for chapter.

sample-aws-agentic-ai-workshop/
├── code/                         # all workshop code, run the commands from here
│   ├── 00-setup/                 # environment setup, uv project, dependencies
│   │   ├── pyproject.toml
│   │   ├── uv.lock
│   │   └── create-uv-env.sh
│   ├── 01-single-agent/
│   │   ├── labs/                 # you write these
│   │   └── completed/            # reference implementation
│   ├── 02-multi-agents/
│   ├── 03-chatbot-app/
│   ├── 04-observability/
│   │   └── docker/               # OTel collector + Jaeger
│   ├── 05-agent-memory/
│   ├── 06-agentcore-runtime/
│   ├── 07-agentcore-observability/   # console-only chapter, no code
│   └── 08-kiro-dev/
│       └── .kiro/                # steering rules and MCP config
└── workshop/                     # all lab guides
    ├── en/                       # English guides, one folder per chapter
    │   ├── 00-setup/README.md
    │   ├── 01-single-agent/README.md
    │   └── ...
    ├── ko/                       # Korean guides, same folder names
    │   ├── 00-setup/README.md
    │   ├── 01-single-agent/README.md
    │   └── ...
    └── images/                   # screenshots referenced by the guides

Every chapter follows the same shape:

code/NN-chapter/
├── labs/                    # empty files, you write the code here
└── completed/               # reference answers, run these if you get stuck

workshop/en/NN-chapter/README.md # English lab guide
workshop/ko/NN-chapter/README.md # Korean lab guide

---

💰 Cost and cleanup

The labs call Bedrock models on demand, and several chapters create AWS resources that bill for as long as they exist:

Chapter	Standing resources
01	Bedrock Knowledge Base, OpenSearch Serverless collection, S3 bucket
05	AgentCore Memory resource
06	AgentCore Runtime, ECR repository, IAM execution role, CloudWatch log groups
07	CloudWatch Transaction Search ingestion, trace and log retention

Warning

Each chapter has its own Cleanup section. Work through them when you are done, especially chapters 01, 05, and 06. An OpenSearch Serverless collection in particular bills continuously whether or not you query it.

---

🐛 Troubleshooting

These are the failures that come up across chapters. Each chapter also has its own Troubleshooting section for problems specific to it: 00, 01, 02, 03, 04, 05, 06.

Symptom	Cause and fix
uv: command not found	The installer puts the binary in ~/.local/bin. Open a new shell, or export PATH="$HOME/.local/bin:$PATH".
ModuleNotFoundError for strands or bedrock_agentcore	You are on the system Python. Use uv run --project 00-setup python ..., or activate 00-setup/.venv.
AccessDeniedException when a lab calls a model	Model access is not enabled for that model ID in us-west-2, or your credentials lack bedrock:InvokeModel. Check the model access page.
ValidationException mentioning a region, or model not found	Your default region is not us-west-2. Check with aws configure get region.
Cannot connect to the Docker daemon in chapter 04 or 06	Start Docker Desktop (or Finch / Podman) and confirm with docker info, then re-run.
Memory calls fail right after creating a memory (chapter 05)	A new AgentCore Memory takes 1 to 2 minutes to reach ACTIVE. Wait, then retry. LTM extraction is also asynchronous.
CloudWatch GenAI Observability dashboards are empty (chapter 07)	Transaction Search must be enabled, and the agent from chapter 06 must have been invoked at least once.

---

📚 Additional resources

Official documentation

• Strands Agents documentation
• Strands Agents observability and evaluation
• Multi-agent patterns
• Amazon Bedrock user guide
• Amazon Bedrock AgentCore developer guide
• AgentCore starter toolkit
• Model Context Protocol specification
• AWS MCP servers

Code and samples

• Strands Agents SDK source
• Strands Agents built-in tools
• Strands Agents samples

Security

See CONTRIBUTING for how to report a security issue.

License

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