One supervisor. Every agent reports to it.

The Supervisor is a real LLM with the same memory and tools as the agents it dispatches — but a different prompt and a different scope. Its job is not to write code. Its job is to decide what gets built, by whom, and in what order.

You steer it the way you would steer a tech lead: a goal, a constraint, a correction. It handles the decomposition.

Every agent runs a role — a system prompt plus a tool allow-list. Five ship built in; you can author more as markdown files committed alongside your code.

The Supervisor proposes; you dispose. It shows you its plan before spawning anything, streams every agent's terminal as they work, and asks before anything destructive. Nothing reaches your main branch without a decision from you.

• Plain-language briefs — no config files, no DSL to learn.
• Every agent's live terminal output streamed back to you.
• Approve, redirect, or kill any agent at any point.