Class Central is learner-supported. When you buy through links on our site, we may earn an affiliate commission.

Building Safe and Reliable AI Agents on AWS

Amazon Web Services via Coursera

Go to class Write review

Overview

Google, IBM & Meta Certificates – 40% Off

One plan covers every Professional Certificate on Coursera.

Unlock All Certificates

This course teaches learners how to design AI agents that are trustworthy enough for production use. Using a customer service agent example built on AWS, learners explore the major ways agents fail in real-world environments and the architectural patterns used to reduce those risks. The course covers agent harness design, layered guardrails, steering, tool safety, policy enforcement, evaluations, observability, and cost controls. By the end of the course, learners will be able to explain and apply practical mitigation techniques for deploying safer AI agents with AWS services such as Amazon Bedrock, Strands Agents SDK, Agent Runtime, Agent Gateway, Lambda, and AgentCore policy.

Syllabus

Why Production Agents Need a Safety Harness

This module introduces the central challenge of production AI agents: balancing usefulness with risk. Learners examine real-world failures, the four major failure modes, and the concept of the agent harness as the system layer where safety, reliability, and observability are engineered.

AWS Foundations for a Production Agent

This module establishes the technical foundation of the course by introducing the customer service agent example and the AWS services used to build it. Learners explore the basic agent code structure, how tools are exposed through MCP, and where hosting and orchestration happen in AWS.

Guardrails and Steering for Safer Agent Behavior

This module focuses on layered controls that help agents stay on topic and follow business rules. Learners examine why prompt-only safety is insufficient, then explore Amazon Bedrock Guardrails and Strands steering patterns, including LLM-as-a-judge and deterministic workflow checks.

Safe Tool Use, Policy Enforcement, and Access Control

This module explores how agents interact with tools and why that interaction must be tightly controlled. Learners study risks caused by hallucinated tool inputs, unclear tool design, and context-dependent authorization. They then learn how interceptors, intent-based tool design, and Cedar policies can reduce those risks.

Evaluations, Observability, Cost Controls, and End-to-End Architecture

This module brings together the operational practices needed to run agents reliably over time. Learners explore evaluation strategies that go beyond the happy path, real-time observability, runtime cost limits, and the full end-to-end architecture of the example customer service agent.