Introduction

Agent Swarm is an AI coding system that turns a chat conversation into working software — with structured memory, autonomous agents, and zero context-management overhead.

It sits in the gap between three kinds of tools:

• IDE-bound tools are tied to your editor and a folder. They autocomplete, but they don't manage work — you still track what was requested versus what was built.
• Chat-only tools can discuss code but have no filesystem and no computer, so you copy-paste back and forth.
• Cloud agents have a computer, but it isn't your computer, and the feedback loop is poor.

Agent Swarm gives you a chat interface backed by a real filesystem, autonomous background agents, and a memory system that eliminates the context-window problem.

The core idea

You describe an outcome. A persistent lead agent turns it into work, and a swarm of agents executes it in parallel — each in its own isolated cloud sandbox, each on its own git branch. When the work is done, you review the diff and merge it or open a pull request.

You ──▶ Lead agent ──▶ swarm of agents
                       ├─ planner   (read-only sandbox)
                       ├─ coder     (writes + runs tests)
                       └─ tester    (records video proof)

What makes it different

• Parallelism by default. A planner can read code while a coder mutates it — they're in separate sandboxes, so they never collide.
• Memory that maintains itself. A knowledge base captures architecture and decisions; background agents keep it current.
• Proof, not promises. Implementation tasks produce a Playwright video of the working feature.

Where to start

1. Read the Workflow to see the end-to-end loop.
2. Learn how Isolated sandboxes keep agents from stepping on each other.
3. Understand Threads & branches and Plan & Agent modes.