Plan Mode: Think Before You Act
Plan Mode is one of Cosine’s most powerful features for complex tasks. Instead of jumping straight into execution, the agent first researches, thinks, and builds a detailed plan — then presents it to you for review before doing anything.
What is Plan Mode?
Section titled “What is Plan Mode?”In Plan Mode, the agent is read-only. It can explore your codebase, search the web, read files, and ask you questions — but it cannot create, modify, or delete anything. The output is a structured plan of what it intends to do.
This is valuable because:
- You can catch misunderstandings before they become wasted work.
- The agent builds significantly more context before acting, which leads to better results.
- You can redirect or refine the approach before committing.
How the Token Window Connects to Plan Mode
Section titled “How the Token Window Connects to Plan Mode”The token window percentage shown in the CLI footer represents how much context the agent has built up. A window at 0% means the agent has no accumulated knowledge — it’s a blank slate. A window at 50% means it has absorbed a large amount of context from your folder, previous sessions, and exploration.
Filling up the context window directs the model toward your specific task more precisely. Plan Mode is one of the best ways to build that context deliberately before executing.
Switching to Plan Mode
Section titled “Switching to Plan Mode”Press Shift+Tab to cycle through modes until you reach Plan (shown in purple). Or use the commands menu (/ or Ctrl+P) and select Change Mode. See the Modes reference for a full breakdown of all four modes.
What Happens When the Plan is Ready
Section titled “What Happens When the Plan is Ready”When the agent finishes its planning phase, a confirmation dialog appears with several options:
| Option | What happens |
|---|---|
| Yes, clear context and auto-accept edits | Clears conversation history, keeps the plan, switches to Auto mode |
| Yes, auto-accept edits | Keeps the full conversation, switches to Auto mode |
| Yes, manually approve edits | Keeps the conversation, switches to Normal mode for step-by-step approval |
| Yes, accept and enable Swarm mode | Hands the plan to a swarm of subagents for parallel execution |
| Type feedback | Revise the plan before committing |
| Reject, and exit plan mode | Discard the plan and return to the previous mode |
When to Use Plan Mode
Section titled “When to Use Plan Mode”Plan Mode is most valuable when:
- The task is large or ambiguous and you want to align on approach first.
- You’re working in an unfamiliar codebase or context.
- You want a written record of what will happen before it happens.
- You intend to hand off the plan to Swarm Mode for parallel execution.
For small, well-defined tasks, you can skip Plan Mode and go straight to Auto. But for anything non-trivial, planning first consistently produces better results.
Key Takeaways
Section titled “Key Takeaways”- Plan Mode is read-only — the agent researches and builds a plan without making any changes.
- Building context before execution leads to significantly better outcomes.
- After planning, you can approve, refine, or hand off to Swarm Mode.
- Use the token window percentage as a rough guide to how much context the agent has accumulated.