Summary
Web Dev Cody walks through the architecture of Agent System Swarm, a self-built agentic sandbox platform that connects to GitHub repositories and triggers Claude Code agents in response to pull request comments and custom slash commands.
The platform scales Amazon EC2 instances on demand, running Firecracker microVMs to isolate each agent job — a design motivated by future multi-tenant support. A Lambda function monitors SQS queue depth and adjusts an EC2 Auto Scaling Group accordingly, keeping costs near $0.09 per instance-hour by scaling to zero when idle. Job records are stored in DynamoDB, and the control plane handles authentication before dispatching work to the queue.
A standout engineering detail is the shared memory system injected into every agent’s system prompt, enabling swarm tasks — such as running 20 simultaneous “code cleaner” agents across a repository — to self-coordinate so no two agents duplicate the same refactor. The video includes a live demo of a 20-agent swarm generating pull requests, a malicious code finder agent triggered by PR comments, and a diagram walkthrough of the full control plane. Cody emphasizes that the architecture is intentionally generic: any Claude Code agent definition can be slotted in without rebuilding infrastructure, making it a reusable foundation for a broad class of automated code-review and maintenance workflows.
📺 Source: Web Dev Cody · Published July 08, 2026
🏷️ Format: Hands On Build
