Agent Demos on AWS

This repository is a small monorepo of AI agent examples built on AWS. It currently contains:

• hr-agents/: resume evaluation systems implemented with multiple AWS agent architectures
• restaurant-agent/: a restaurant support agent for menu search and reservations
• harness/: putting an agent inside a deterministic control path (Bedrock AgentCore harness)

The projects are independent. Each subdirectory has its own deployment flow, dependencies, and detailed documentation.

Repository Layout

.
├── hr-agents/
│   ├── README.md
│   ├── stepfunctions-agent/
│   ├── bedrock-agent/
│   ├── strands-lambda-agent/
│   ├── strands-agentcore-agent/
│   └── samples/
├── restaurant-agent/
│   ├── README.md
│   ├── restaurant_agent.py
│   └── test_restaurant_agent.py
└── harness/
    ├── README.md
    ├── configs/                  # one harness config per agent (instructions + output schema)
    ├── shared/                   # harness_lib.py -- the seam helper
    ├── cloudformation/           # AWS::BedrockAgentCore::Harness + the three scenarios
    ├── scenario1-release-review/ # CodePipeline
    ├── scenario2-extract-judge/  # Step Functions
    ├── eventbridge-variants/     # EventBridge advisory agents
    └── diagrams/                 # AWS-icon, draw.io and Excalidraw diagrams

Projects

HR Agents

hr-agents/ contains several approaches to the same resume-processing problem:

• stepfunctions-agent/: orchestrated workflow with AWS Step Functions and Lambda
• bedrock-agent/: Bedrock Agents-based multi-agent implementation
• strands-lambda-agent/: Strands agents running in Lambda
• strands-agentcore-agent/: Strands agents hosted on Bedrock AgentCore Runtime

Typical flow:

1. Upload a job description to S3.
2. Upload a resume to S3.
3. Trigger automated evaluation.
4. Store structured candidate results in DynamoDB.

Start here: hr-agents/README.md

Restaurant Agent

restaurant-agent/ is a conversational restaurant assistant built around AWS Bedrock AgentCore and Strands. It supports:

• menu and dish discovery
• reservation creation, lookup, updates, and cancellation
• DynamoDB-backed persistence for reservation records

Start here: restaurant-agent/README.md

Harness — agents inside a deterministic control path

harness/ demonstrates how to put a non-deterministic agent inside a deterministic control path safely, using the Amazon Bedrock AgentCore managed harness (native AWS::BedrockAgentCore::Harness in CloudFormation) with Amazon Nova 2 Lite.

A surrounding AWS service hands the agent one well-scoped job. The agent does the fuzzy part and emits structured JSON validated against a schema. The AWS service owns control flow, retries, and any irreversible action. The agent always fails closed.

Same harness pattern, three entry points, one boundary rule:

Entry point	Agent	Who acts on the verdict
CodePipeline action	release-review → pass / concerns / block	the pipeline (Lambda glue)
Step Functions task	extract-and-judge → confidence + flags	the state machine (Choice state)
EventBridge rules	security-responder, deploy-annotator (advisory)	a human

Deployed with plain CloudFormation (no CDK, no custom resource). All three scenarios are verified end to end against a live account.

Start here: harness/README.md

Prerequisites

Most examples assume:

• AWS CLI configured for an account with the required permissions
• Python 3.10+ or 3.11+
• access to Amazon Bedrock in the target region
• AWS SAM CLI for the SAM-based deployments

Some projects also require Bedrock AgentCore or Strands-specific tooling. Check the README in the relevant subdirectory before deploying.

Quick Start

Choose one project and follow its local README:

cd hr-agents

or

cd restaurant-agent

Notes

• This repo is organized as examples, not as a single deployable application.
• Infrastructure, scripts, and sample data live with each project that uses them.
• If you want to compare agent architectures for the same use case, start with hr-agents/.