Create a task

Describe what you want in plain language; an agent does it in an isolated workspace.

A task is one unit of agent work: a prompt, an agent, and a workspace. The agent works in its own branch; you review and merge.

The New Task composer — one sentence is enough; everything else is optional.

The quick path

1. Open a project and the ⚡ Task tab

   On the project page, the “Start work” control has three tabs: ⚡ Task, 🐝 Pipeline, and 🎨 Design.

2. Describe it in one sentence

   e.g. “Add a contact form to the pricing page.” Example chips can fill the prompt for you.

3. Accept the Terms

   Tick the Terms & base-image acknowledgment (see Terms).

4. Start 🐝

   The agent provisions a workspace, clones your repo, branches, and starts. You’re taken to the task screen.

That’s all most tasks need — everything else is optional under Advanced options.

Advanced options

Advanced options: agent, model, Web IDE, resources & add-ons, and race variants.

Expand “Advanced options — safe to ignore” for:

• Task name — optional; derived from your prompt if blank.
• Experience (Claude only) — Code (edit the repo) or Cowork (a general assistant). Design has its own tab.
• Coding agent — Claude Code, Codex, Gemini, or Kiro (limited to your org’s allowlist). See Agents.
• Model — appears when your org allows more than one model for that agent.
• Web IDE — open the task in a full browser VS Code (premium; org-gated). See Web IDE.
• Resources & add-ons — allocate extra CPU/RAM from your org’s purchased pool (fine-grained, free allocation), and optionally enable backup.
• Race variants (Code mode) — run the prompt across several agent/model variants at once and merge the winner. See Multiple models.

A live usage indicator shows how many workspaces (and IDE seats) you’re using against your plan’s cap, and blocks creation if you’re at the ceiling.

Just need a shell?

The project header’s “Open console” starts an interactive shell in the cloned repo (a Claude-agent task with shell mode) — handy for poking around without launching a full agent run.

What happens on submit

The task is created against your project, branching from the default base branch. If you enabled the Web IDE, the task gets a heavier sandbox (1 vCPU / 1 GiB / 1 GiB). Extra resources are allocated from your pool and clamped to what’s available. Then you land on the task screen to watch, review, and merge.