Plan mode

Plan mode puts a checkpoint between an instruction and its execution. Instead of acting on your request immediately, the agent first works out its approach, shows it to you, and waits for a yes. Use it whenever you want to see the plan before the agent commits to it.

#How it works

Turn plan mode on for a channel with the !plan command. While it's on, every task runs through the same flow:

1. Think. The agent reasons through what you asked before doing anything.
2. Propose. It posts a plan card in chat describing what it intends to do — the approach it's going to take.
3. Wait. Nothing executes. No commands run, no files change, until you act on the card.
4. Approve. Once you approve the plan card, the agent carries out the plan.

Toggle it back off with !plan again, and the channel returns to acting on your messages directly.

#When to use it

Plan mode is worth the extra step for work where the approach matters as much as the result:

• Risky changes — anything touching production, credentials or shared state.
• Large tasks — a multi-file refactor or migration where you want to confirm the shape before it starts.
• Irreversible actions — deletes, deploys, or anything hard to undo once it's run.

For quick, low-stakes work it's usually overhead you don't need — that's why it's a per-channel toggle rather than an always-on mode. Turn it on for the channel where the stakes are high, leave it off where they aren't.

Plan mode pairs naturally with the approval buttons that surface on individual sensitive tool calls: one reviews the whole approach up front, the other gates specific actions as they happen. Together they keep an agent with real machine access on a short, visible leash.

#Where to go next

• Chat commands — the full set of ! commands, including !plan.
• Agents & your team — how agents run tasks once a plan is approved.
• Scheduling & automation — hand off longer-running work once you trust the approach.