Enterprise AI Agents on AWS

Lab 2a: deploy an ML model

• Deploy a pre-trained model to a SageMaker AI real-time endpoint
• Test it
• Autoscaling
• Serverless inference option

Lab 2b: use foundation models

• Bedrock Converse API
• On-demand and provisioned throughput
• Cross-Region inference
• Open models in SageMaker JumpStart

Customize and control

• Bedrock model customization (demo)
• Prompt caching and token cost
• Quotas and throttling

Outcome:

• One ML model live on an endpoint and foundation models called from code.

Prompting and structured output

• System prompts and inference parameters
• Structured output
• Tool use with the Converse API

Lab 3a: RAG with Bedrock Knowledge Bases

• Amazon S3 data source
• Chunking and embeddings
• Vector store options
• Retrieve-and-generate with citations

Lab 3b: guardrails and evaluation

• Bedrock Guardrails
• Content filters
• Denied topics
• Sensitive-data redaction
• Grounding checks
• Bedrock evaluations

Outcome:

• A RAG application over your own documents, with guardrails on and evaluation scores.

Agent design

• Roles and instructions
• Tool schemas
• Conversation state and memory

Lab 4a: build an agent with Bedrock Agents

• Instructions and action groups
• AWS Lambda functions as tools
• Attaching a knowledge base
• Session memory

Lab 4b: build an agent with Strands Agents

• Strands Agents SDK in Python
• Custom tools
• MCP servers
• AgentCore Gateway to expose APIs and Lambda functions as tools

Outcome:

• Two working agents with tools, knowledge and memory: one in Bedrock Agents and one coded with Strands.

Orchestration patterns

• Supervisor with collaborator agents
• Agents as tools
• Swarm and graph patterns
• When one agent is enough

Lab 5: build a multi-agent system

• Supervisor and specialist agents in Strands
• Shared context and handoffs
• Human-in-the-loop approval

Agent-to-agent communication

• The A2A protocol
• Calling a remote agent
• Bringing a LangGraph or CrewAI agent onto AgentCore

Outcome:

• A multi-agent system with a supervisor, specialist agents, human approval and an A2A connection.

Lab 6a: deploy agents

• Package and deploy to AgentCore Runtime
• AgentCore Identity and Memory
• Infrastructure as code and CI/CD (walkthrough)

Lab 6b: observability and evaluation

• AgentCore Observability with Amazon CloudWatch
• Traces, token use and latency
• Agent evaluation and error analysis

Governance, cost and wrap-up

• Guardrails on agents
• IAM and tool-access policies
• Cost controls
• Final demo
• Next steps toward AIP-C01

Outcome:

• The multi-agent system running on AWS with tracing, guardrails and cost controls, demonstrated end to end.