Planning mode
The Agent writes a plan first and only touches the project once it is approved. When that pays off and when it is just a delay.
In planning mode the Agent does not rewrite the project. It may read and it may think; at the end it writes a plan you read and approve, or send back. It is the one moment when it makes sense to wait for the Agent — before it starts writing, not afterwards.
How it is turned on
In the Conversation, under the + button, is the “Planning mode” switch. It stays on until you turn it off, and it is visible — the button is highlighted, so you will not hand over work thinking you are planning.
The Prompt field asks differently as well: “What should the Agent plan?” instead of the usual invitation.
Switching the mode calls for a restart of the Agent on the Server. That happens by itself and is a matter of a second or two — but that is exactly why the mode is not switched for every Prompt.
What cannot happen while planning
The Agent runs with writing forbidden. It cannot write a file, it cannot run a command that changes anything. It is not a promise in the task that it could overlook — it is the setting it was started with.
The plan then does not appear as an ordinary reply. The Agent submits it as a request for approval and you see the whole plan, not its beginning — for other actions it is enough to know what will run, here that text itself is the content of the decision.
Approve, or send back
With a plan you have two options and both are fine:
- Approve the plan
- The Agent finishes the turn and waits. A button appears under the Conversation: Run the plan — one tap instead of writing “go ahead then”. Only that sets it to work, already in the ordinary mode.
- Turn it down and write what is wrong
- You write what does not fit in the plan and the Agent reworks it — still planning, still without writing. This is the cheap way to steer it somewhere: fixing a plan is an order of magnitude faster than fixing finished code.
When to use it
- For a task you are not sure you described well. A plan is the fastest check that you understood each other.
- In an unfamiliar project. Before the Agent rewrites anything, you want to see which files it will touch.
- For a change across the project — a refactor, a rename, a migration. There a misunderstood task shows up in fifty places.
- When you work without approvals. Bypass and planning are a pair that makes sense: you go through the big tasks with a plan and the rest then runs without interruptions.
When not to use it
For small tasks a plan is a delay. “Fix the typo in the README” does not need a three-point deliberation, and reading it costs more than the fix.
It also does not suit cases where the Agent has to try something first to know what to propose — debugging, say, where the cause is found by running a test. While planning it may not run anything, so it would be planning blind.
A plan the Agent asks for itself
Sometimes a plan arrives even though you did not turn planning on — the Agent decides on its own that the task calls for deliberation. It behaves the same: you read it, approve it and “Run the plan” appears.
Stuck somewhere other than what is here? Write to support@coden-app.com.