04-ecs-agent-orchestrator

ECS Agent Orchestrator

Long-running agent on ECS Fargate that orchestrates Lambda tools using ReAct pattern with streaming responses.

What This Demonstrates

• ReAct (Reasoning + Acting) agent pattern implementation
• ECS Fargate for long-running containerized workloads
• Lambda function orchestration from ECS
• Server-Sent Events (SSE) for streaming responses
• Multi-step reasoning and tool usage loops

Prerequisites

• AWS ECS cluster or Docker for local development
• Deployed Lambda AI tool from Example 03
• Anthropic API key
• AWS credentials with Lambda invoke permissions

Quick Start

npm install
docker build -t ecs-agent . && docker run -p 3000:3000 ecs-agent
curl -X POST http://localhost:3000/agent/process -d '{"task":"Analyze customer feedback and suggest improvements"}'

How It Works

The agent follows a ReAct loop for complex reasoning and action:

1. Reasoning: Agent analyzes the current state and decides next action
2. Acting: Calls appropriate Lambda tools (summarize, classify, extract)
3. Observing: Processes tool results and updates internal state
4. Repeating: Continues until task completion or iteration limit

User Request -> ReAct Loop -> Stream Progress -> Final Result
     |              |              |              |
     v              v              v              v
  Parse Task -> Reason/Act -> Lambda Tools -> Synthesize

The agent streams its reasoning process and intermediate results via Server-Sent Events, allowing real-time visibility into multi-step workflows.

Limitations

While this approach handles complex workflows Lambda cannot:

• Container overhead increases costs vs serverless
• Always-running resources vs event-driven Lambda
• Complex scaling decisions vs automatic Lambda scaling
• Infrastructure management vs managed Lambda runtime

Next Steps

See Example 05 (streaming-chat) for user-facing chat interfaces, or Example 06 (cost-tracking) for monitoring multi-step agent costs.