Agent Teams

Agent Teams let one coding agent session spawn parallel teammate agents. Super Loops run an agent iteratively on one task until it is complete.

Agent Teams and Super Loops are two ways to get more out of a coding agent than a single linear session. Agent Teams let one session split its work across parallel teammate agents; a Super Loop runs one agent over and over on a single task until the task is done. Running fan-out work across several agents is covered on the agent orchestration page.

Agent Teams

Agent Teams let a single coding agent session spawn teammate agents that work in parallel. This suits work that splits into independent subtasks, at the cost of using more tokens.

To turn it on:

1. Go to Settings and open the Claude Agent panel.

2. Enable Agent Teams (Experimental).

3. Start a session. When the agent decides subtasks can run in parallel, it spawns teammates automatically.

4. Each teammate shows a distinct progress indicator in the transcript.

Good to know:

• Teammate output stays contained; it does not leak into the main transcript.

• The main session waits for all teammates to finish before completing.

Super Loops

A Super Loop runs an autonomous agent iteratively until a task is complete. Each iteration starts with fresh context while progress persists via files, so the loop can chip away at a large task without ever filling up one session's context window. Super Loops require a git repository; a dedicated worktree is created automatically for each loop, so its changes stay off your main branch.

To start one:

1. Open Settings > Application > Agent Features and enable Super Loops.

2. Return to Agent mode, open the menu beside New session, and choose New Super Loop. This option is disabled until the project is a git repository.

3. Describe the task. The description is saved to .superloop/task.md in the new worktree.

4. Pick the model and the maximum number of iterations, then create the loop.

The agent then works in iterations, reading its own progress files at the start of each one, until the task is complete or the iteration limit is reached.