Permissions

Agents can read files, edit code, and run commands. The permission mode decides how much of that happens without asking you first.

The permission selector

It lives in the composer, next to the model selector, and applies per conversation. Modes range from cautious — the agent asks before each meaningful action — to autonomous build modes where it proceeds and reports.

Choosing a mode

• Starting out: keep the default. You will see each proposed action and approve it, which is the fastest way to learn what an agent actually does.
• Trusted, repetitive work: raise the autonomy so the agent stops asking for actions you always approve.
• Teams: the leader’s permission choices propagate to teammates, so a team behaves consistently.

What stays in your hands regardless

• Attaching the workspace: agents work in the folder you give them.
• The conversation transcript records the agent’s actions and tool calls, so autonomy never means invisibility.

Related: Agent teams →