Running agents in parallel
Atelier is a canvas, and agents are panels on it. This guide covers launching them, reading their status at a glance, and the tools you'll want around them.
Launching an agent
Add an agent panel and pick your provider — Claude Code, Codex, or any CLI agent already on your PATH. Atelier is the workspace, not the model: it finds the tools you already have installed and runs them in your project, so your existing plan keeps working exactly as it does in a terminal.
There is no limit on how many run at once. A typical setup is three to five agents side by side, each on its own task — and if they work on the same repository, you'll want each in its own isolated worktree so they never step on each other.
Reading the canvas
Every agent panel carries a live status: working, waiting on you, or done. The state comes from the agents' own lifecycle signals rather than from guessing at terminal output, so “done” means the agent actually finished its turn — not that it merely went quiet.
Zoom out and the canvas becomes an overview of everything in flight; zoom in to work with one agent at full size. The point is that five agents stop being five terminal tabs you cycle through blindly, and become one picture you can read.
When an agent needs you
The moment an agent stops to ask something, its panel says so. You answer where the question is, and the agent continues. Waiting is the state that costs you money and time — Atelier exists to make it impossible to miss.
The tools around your agents
Everything an agent's work needs lives on the same canvas: terminals that open in the right checkout, a browser panel for localhost previews, an editor for quick changes, and notes for the plan you're carrying in your head. Group a task's panels into a Work Area and the whole group moves, and persists, together — reopen Atelier and everything is where you left it.
Privacy, stated plainly
Your prompts, code and transcripts never leave your Mac. What Atelier keeps, it keeps locally. The privacy statement lists everything that does leave, field by field — it is a short list, and none of it is your work.
Next: isolate parallel agents with git worktrees, or go back to the quick start.