Permission modes
Permission modes control how freely an agent can act in a task: whether it only plans, asks for approval before acting, edits files automatically, or works with full access.
The selector shows the selected agent's own modes, as that agent defines them, so the list changes with the agent. You can select a mode on the New Task page and change it at any point while the task runs.
Select a permission mode
Under the task input, click the permission mode selector, next to the model selector.
Select the mode that matches your intent. For example, Claude Agent offers:
Plan – analyze the task and propose a plan without changing code. See Plan before the agent changes code.
Ask Permission – require approval before making changes or running commands.
Auto-Edit – allow the agent to edit files automatically, while still keeping more limits than full access.
Full Access – allow the agent to work with the fewest restrictions available for that agent. Use this mode only if you trust the project and understand the risks.
OpenAI Codex offers Chat, Agent, and Full Access, with no planning mode of its own.
When the agent asks for permission, answer in the card that appears in the chat. While a task waits for your answer, the editor shows a Task needs your input notification.
Plan before the agent changes code
If the agent offers a planning mode, such as Plan in Claude Agent, select it when you want to agree on the approach before any code changes. In this mode, the agent reads the code, asks clarifying questions, and answers with a plan in the chat instead of editing files.
Planning is useful for tasks with several conditions or constraints, changes that span several files or modules, and tasks where you want to confirm the scope first. For small, obvious changes, skip it.
When the plan suits you, switch to a mode that lets the agent edit files, such as Auto-Edit, and ask it to implement the plan. If it doesn't, reply in the chat and let the agent revise it.