> ## Documentation Index
> Fetch the complete documentation index at: https://www.rumus.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Plan mode

> How the agent automatically drafts a checklist for multi-step jobs and works through it step by step.

For non-trivial tasks the agent doesn't just dive in — it drafts a **plan** first, walks through it step by step, and updates the checklist as it goes. Plan mode is **automatic**: the agent decides when a request is big enough to warrant planning. There's no toggle to flip.

## When plan mode kicks in

The agent enters plan mode when your request implies multiple steps that need to be sequenced. Some signals:

* "Set up …" — installs, configurations, multi-stage rollouts
* "Migrate …" — anything with reversal cost
* "Investigate why … and fix it" — diagnostic-then-action work
* Anything where running the wrong step first would waste time or cause harm

Quick one-shots ("show me the disk usage", "summarize this log") don't trigger planning — there's no sequence to plan.

## What you see in the sidebar

When the agent decides to plan, a **Plan block** renders in the conversation. It looks like a checklist with a goal at the top and a numbered list of steps below.

Each step has a status icon:

| Icon        | Meaning                                      |
| ----------- | -------------------------------------------- |
| ` ` (blank) | Pending — not started yet                    |
| `→`         | In progress — currently working on this step |
| `✓`         | Completed                                    |
| `✗`         | Failed — the step didn't succeed             |

Step format: `[icon] N. Title — Note`. The note is optional and shows context (e.g. *"used `apt` because the host runs Ubuntu"*).

## How execution works

<Steps>
  <Step title="Plan drafted">
    The agent posts the plan with all steps **pending**. It pauses here for your review — you can edit the plan or accept it.
  </Step>

  <Step title="Approve to start">
    Click approve. The agent updates step 1 to **in progress** and starts working.
  </Step>

  <Step title="Step-by-step execution">
    As the agent finishes each step, the icon flips to **completed** and the next step becomes **in progress**. If a step fails, the agent marks it **failed** and decides whether to retry, skip, or stop.
  </Step>

  <Step title="Plan updates">
    Subsequent updates to the plan (new steps discovered mid-task, reordering) happen silently — the same Plan block updates in place.
  </Step>

  <Step title="Done">
    When every step is completed (or failed and skipped), the agent posts a final summary message under the Plan block.
  </Step>
</Steps>

## Editing the plan before approval

The first time the plan is posted (before you've approved), you can revise it:

* **Edit a step** — refine the description.
* **Reorder steps** — change the order of operations.
* **Remove steps** — drop ones that aren't needed.
* **Add steps** — insert your own.

This is the moment to catch a missed prerequisite or a step that targets the wrong host. Once you've approved, edits to the plan come from the agent.

## Cancelling mid-plan

Hit **Stop generating** at any time. The agent finishes the current tool call (it doesn't kill in-flight commands) and stops. The plan stays visible in the conversation — completed steps are still marked completed, the in-progress one keeps its icon, and pending steps remain unstarted.

You can come back to the conversation later and ask the agent to resume from where it left off.

## When NOT to use plan mode

You can always *avoid* plan mode by giving the agent a single-step task. If the agent enters plan mode for something that's actually a one-shot, say so ("just run X") and it'll skip planning next time.

Conversely, you can encourage plan mode for jobs that look ambiguous by being explicit: "Plan first, then execute step by step."

## Tips

* **Keep the screen visible during the first run of a new plan.** The plan block updates inline — watching it is the fastest way to learn how the agent reasons about your environment.
* **Tight skills + plan mode is a cheat code.** A well-defined [skill](/docs/ai/rules-skills) named "deploy-staging" turns into a tidy single step in any plan that needs to deploy.
* **For risky multi-stage work, narrow the scope.** Big plans are slower and harder to review than small ones. "Set up monitoring on prod-web-1" beats "set up monitoring on prod."

## Next steps

<CardGroup cols={2}>
  <Card title="Agentic execution" icon="robot" href="/docs/ai/agentic">
    The mechanics of how each step actually runs.
  </Card>

  <Card title="Sub-agents" icon="diagram-project" href="/docs/ai/sub-agents">
    Independent steps the agent fans out to parallel workers.
  </Card>
</CardGroup>