Workshop Outline

This workshop is structured as a progressive, hands-on journey. You’ll start by building a simple conversational agent and incrementally add capabilities (tools, memory, multi-agent orchestration, safety guardrails, and production deployment) until you have a fully functional SupportBot running on AWS. Each module builds directly on the previous one, so by the end you’ll understand how every piece fits together.

Time	Module	Topic	Key Activity
0:00	Setup & Intro	Environment + Concepts	Run verify_setup.py + presentation
0:20	Module 1	Your First Agent	Build basic FAQ agent
0:45	Module 2	Custom Tools & MCP	Add tools + build MCP server
1:20	Module 3	Memory & Context	Add persistent memory
1:50	Module 4	Multi-Agent Patterns	Build triage + specialists
2:20	Module 5	Evals, Safety & Observability	Run eval suite + guardrails
2:50	Module 6	Deploy to AWS	Deploy to AgentCore
3:15	Wrap-up	Review & Next Steps	Q&A

---

Learning Objectives by Module

Module 0: Environment Setup & Intro (20 min)

• Verify Python 3.12+, AWS CLI, and Bedrock access
• Install Strands Agents SDK and dependencies
• Understand the project structure

Module 1: Your First Agent (25 min)

• Create a Strands Agent with a system prompt
• Understand the agentic loop (reason → respond)
• Test conversational interactions
• Key concept: Agent = Model + System Prompt + Tools

Module 2: Custom Tools & MCP (35 min)

• Build tools with the @tool decorator
• Understand tool schemas from docstrings and type hints
• Build an MCP server with FastMCP
• Connect agent to MCP server via stdio transport
• Key concept: Tools extend agents from “answering” to “doing”

Module 3: Memory & Context Management (30 min)

• Implement persistent customer memory
• Learn hierarchical context (L1/L2/L3)
• Understand token optimization strategies
• Explore the Skills.md pattern
• Key concept: Load context on demand, not all at once

Module 4: Multi-Agent Patterns (30 min)

• Build specialist agents (billing, technical, returns)
• Implement triage agent with Agents-as-Tools pattern
• Understand Swarm, Graph, and Workflow patterns
• Key concept: Multi-agent enables specialization

Module 5: Evals, Safety & Observability (30 min)

• Build an evaluation suite with test cases
• Implement input/output safety guardrails
• Use LLM-as-Judge for quality assessment
• Set up OpenTelemetry tracing
• Key concept: Agent evals are non-deterministic. Score, don’t pass/fail

Module 6: Deploy to AWS AgentCore (25 min)

• Wrap agent with BedrockAgentCoreApp
• Deploy with AgentCore starter toolkit
• Test the deployed agent
• Key concept: AgentCore = managed agent infrastructure

---

Prerequisites

• Python 3.12+
• AWS account with Bedrock access (Claude Sonnet enabled)
• Basic Python knowledge
• Familiarity with the command line
• No prior AI/ML experience required

What Attendees Walk Away With

1. A working multi-agent customer support system
2. Understanding of agentic AI architecture patterns
3. Hands-on experience with Strands Agents SDK
4. Knowledge of MCP, memory, evals, safety, and observability
5. A deployed agent on AWS AgentCore
6. Access to all workshop materials and code