Modes and when to use them
Plan mode, edit mode, auto-approve: what each one changes, and how to move between them on purpose.
Most harnesses ship an agent mode: a preset bundling a permission posture with behavioural instructions. Names differ; the shapes are consistent.
The four shapes
| Shape | What it does | Use when |
|---|---|---|
| Ask / chat | Reads and answers. No writes. | Exploring, understanding, deciding. Module 04’s exploration work lives here. |
| Plan | Reads and proposes a written plan. No writes until you accept. | Anything touching more than one file. The default for real work. |
| Edit | Writes files, asks before shell commands. | Executing an agreed plan while you watch. |
| Auto / full-auto | Writes and runs without asking. | Well-specified work in a sandbox. Never on an unsandboxed machine with live credentials. |
The escalation habit
Mode should follow confidence, and confidence should follow evidence. A session that starts in ask mode, moves to plan once you understand the shape, then to edit once the plan is agreed, is doing something useful: each step down the ladder is gated by something you now know.
ask -> "how does the retry logic work today?" (read only, cheap) plan -> "propose how to add exponential backoff" (still no writes) edit -> "implement the plan, run the check command" (writes, watched)
Going straight to edit is fine for work you have done a hundred times. It is the default that gets people into trouble on unfamiliar code.
Plan mode is underused
The most common upgrade in someone's practice is simply using plan mode more. It costs one extra turn and it front-loads every correction to the point where corrections are cheap. It also produces an artifact — the plan — that you can save, edit, and carry into a fresh session, which is the basis of the whole decomposition approach in Module 06.
Watch out
Auto mode plus a long session is the worst combination available: unsupervised work, accumulating context, silent degradation. If you want unattended runs — and they are genuinely useful — give them a tight spec and a sandbox, and keep them short. Module 06 covers this properly.
Try it
Take the medium task from your list and run the full escalation: ask, then plan, then edit, clearing nothing in between. Note where the plan needed correcting. That correction is what you would have paid for in code otherwise.
Takeaways
- Match mode to confidence and escalate as evidence accumulates.
- Plan mode is the highest-return habit change available to most people.
- The plan is a reusable artifact, not just a checkpoint.
What makes auto mode plus a long session particularly dangerous?
The two failure sources compound and neither is visible. Context degrades silently while nobody is reviewing, so quality drops exactly when supervision is absent. Unattended work needs to be short, tightly specified, and contained.
A course by Pieter Zandbergen