FitWhen to use it — and when not
Use it when
- Multi-step tasks longer than a minute
- Tasks that touch external systems or cost money
- Ambiguous requests where the agent made assumptions
Skip it when
- One-step tasks
- Routine automations the user already approved as a template
AnatomyThe parts of the pattern
- StepsShort, verb-first steps in order.
- Risk labelsRead-only, writes data, sends externally.
- Edit controlsToggle, reorder, rewrite steps.
- EstimateRough time and cost.
- RunExecute the plan as edited.
GuidelinesDo & don’t
Do
- Label which steps change things outside the product.
- Let users remove a step without rewriting the request.
- Show assumptions the agent made, as editable lines.
Don’t
- Show a 20-step plan for a 2-step job.
- Ignore the edited plan once execution starts.
- Present the plan as a wall of prose.
In the wildReal-world examples
Claude Code (plan mode)CursorGitHub Copilot WorkspaceDevin
Products named for reference only — no affiliation, and the demo above is an original illustration, not a copy of their UI.
For engineersImplementation notes
- Have the planner return steps as structured data (id, action, tool, risk) and feed the edited list back as the execution contract.
- Re-plan only the remaining steps when one fails; keep completed step ids stable.
- Estimate cost from step types and historical runs, not from the model's own guess.