Skip to content

Modes: Plan, Auto, Effort

The same agent can be a careful architect or a fast pair of hands. You choose, per task, with three dials.

Plan mode

Press Shift+Tab to cycle into plan mode: the agent can read, search, and think, but not modify anything. It explores your code, then comes back with a plan you approve, edit, or reject before any file changes.

Use it when:

  • the task touches many files or a codebase the agent hasn't seen;
  • you're not sure what you want yet — the plan is a cheap draft of the work;
  • being wrong is expensive (refactors, data migrations, anything with rm).
(plan mode)
> Our analysis notebook has grown to 2000 lines. Propose a split into a
> proper package with tests, keeping the notebook as a thin driver.

The plan you get back is a negotiation document. Push back on it. Editing a plan costs seconds; un-confusing a half-refactored repo costs an afternoon.

Auto-accept

The opposite dial: stop confirming every edit and command, and let the agent run. Shift+Tab cycles to accept edits; the --dangerously-skip-permissions flag (or a sandboxed environment) goes further.

The rule of thumb:

Auto-accept exactly when there's a feedback loop that catches mistakes mechanically — tests, a compiler, a metric. Confirmation by human is for work that has no such loop.

This is why feedback loops and auto-accept are really one pattern: the loop is what makes autonomy safe.

Effort & model choice

Reasoning effort and model size are a budget knob, not a quality cult.

  • Small/fast models, low effort — mechanical transformations, renames, boilerplate, "apply this pattern to these 12 files."
  • Big models, high effort ("think hard") — architecture, debugging something genuinely confusing, deriving the right equations once before generating code from them.

From Inar's intro you know the economics: input tokens are cheap and cacheable, output tokens and reasoning are what you pay for. Spend reasoning where being wrong is expensive, and stream boilerplate from the cheap end of the dial.

The combination that works

Plan mode with a big model to decide what to do, then auto-accept with a fast model to do it, inside a feedback loop that verifies it. That's the workhorse configuration for research code.