Architecting Enterprise Agentic Systems on AWS

Description

This two-day, instructor-led course teaches experienced agent builders how to architect agentic systems that an enterprise can trust, afford, audit, and operate. It starts from one rule: build the simplest system that meets the requirement, and add agents only when a measured result says they earn their cost. Participants build a single-agent baseline and an evaluation harness first, then judge every later design, including multi-agent ones, against that baseline on quality, token cost, and latency. Participants engineer the context an agent sees. They choose a coordination pattern and shared memory only where context isolation, parallel work, or specialization demands it. They connect the system privately to enterprise data and APIs, produce the audit evidence a compliance team asks for, control its cost and resilience, and release new versions safely under continuous evaluation. The capstone is a production-readiness review of a complete system, defended against a rubric. About two-thirds of the course is hands-on, with less step-by-step guidance than LF-20302: labs state the requirement and the acceptance test, and participants choose the design. Every lab runs in a hosted AWS sandbox and uses foundation models on Amazon Bedrock, selected per lab through a parameter.

Who This Course Is For

• Solutions architects and principal engineers who design agent platforms for an enterprise
• Senior software and ML engineers who have shipped an agent to production and now own several
• Platform, security, and SRE leads who must approve, isolate, audit, and operate agentic systems
• Technical leads deciding whether a workload needs one agent or several

What You Will Learn

• Select an agent architecture for a workload by measuring candidate designs against a single-agent baseline, and adopt multi-agent coordination and shared memory only where context isolation, parallel work, or specialization justifies their cost.
• Engineer an agent's context and cost profile, managing what the model sees, what is cached, and what is remembered, and keeping the system performant and resilient.
• Secure and audit an agentic system at the network, data, and trust-boundary levels, including private connectivity to enterprise systems, defenses against injected instructions, human approval of irreversible actions, and audit evidence.
• Operate and release an agentic system under end-to-end observability and continuous evaluation, and defend its production readiness with evidence.

Prerequisites

• LF-20303 _AI Agent Fundamentals on AWS_, or equivalent experience building a tool-using agent
• LF-20302 _Engineering Production AI Agents with Amazon Bedrock AgentCore_, or equivalent production experience with AgentCore Runtime, Gateway, Memory, Identity, Policy, and Observability; this course does not re-teach those services
• Working Python, including async code
• Comfort with Amazon VPC basics (subnets, route tables, and security groups), IAM policies, and AWS CloudFormation
• A modern web browser and a reliable internet connection; the lab sandbox is provided

Delivery Options

• Live, instructor-led
• Virtual instructor-led