Agent mode
Also called Mode
A named preset that bundles a permission mode with behavioural instructions, switchable in the middle of a session.
An agent mode is a preset the harness offers that changes two things at once: the permission mode (what can run without asking) and a slice of the system prompt (how the agent is told to behave). 'Plan' modes forbid edits and instruct the model to investigate and propose. 'Accept edits' modes let file changes flow but still gate commands. 'Ask' or 'chat' modes turn the agent into a consultant that touches nothing. You can flip between them inside one session without losing context.
Modes are worth understanding because the same model behaves noticeably differently under each, and people blame the model. An agent that 'refuses to make changes' is often just sitting in a read-only mode. An agent that 'went rogue and edited twelve files' was in an auto-accept mode you toggled earlier and forgot about. The mode indicator in the UI is the first thing to check when behaviour seems off.
A good rhythm uses modes deliberately: start a task in plan mode to align on the approach, switch to an editing mode to execute, and drop back to a read-only mode to review. Treat the mode switch as part of the steering vocabulary, alongside interrupting and re-prompting.
- Claude CodeShift+Tab cycles default, accept edits and plan; the current mode is shown next to the input.
- CursorAgent, Ask and Plan are selectable in the chat panel and set what the assistant may touch.
- CodexThe mode is expressed as an approval policy plus sandbox settings rather than a single named toggle.
“It wrote a beautiful plan and then just stopped.”
“You're in plan mode. That's the deal: it plans, you approve, then you switch modes and it builds.”
“Ah. Shift+Tab, then.”
Course 01 puts every one of these terms to work: you install Claude Code, run the loop, and ship a real project — permission modes, compaction, hooks and all.
Start Course 01 →