Introducing Agent Swarm
Why we built a chat that turns conversations into shipped software — and how a swarm of isolated agents makes it work.
Every AI coding tool today forces a trade-off. IDE tools are glued to your editor and a folder. Chat tools can talk about code but can't run it. Cloud agents have a computer, but it isn't yours, and the feedback loop is painful.
We wanted something different: a chat interface with a real filesystem, autonomous background agents, and a memory system that makes the context window a non-issue.
A conversation, not a copilot
You don't open a project to talk to Agent Swarm — you describe an outcome. A persistent lead agent turns that into work and hands pieces to a swarm of agents that run in parallel.
The key insight is that parallelism needs isolation. A planner that reads code shouldn't share a filesystem with a coder that's rewriting it. So every agent gets its own cloud sandbox, branched from main.
Proof, not vibes
"It works" should mean there's a recording that proves it works. Every implementation task produces a Playwright video of the feature actually running, with acceptance criteria encoded as tests that run on every change.
What's next
This is the start. We're deepening the knowledge system, expanding automations, and making review even tighter. If you want to build with a swarm, start a project — the free tier is enough to feel the difference.