Chatting: Listen, Plan, Go
By default, you type a message and the agent gets to work. Three optional modes sit beside the input box and change that. They're mutually exclusive — turning one on turns the others off — and each applies until you turn it off.
Who can do what
| Task | User | Agent owner | Admin |
|---|---|---|---|
| Use any of the three modes | ✅ | ✅ | ✅ |
Modes are a property of how you are chatting, not of the agent, so anyone who can talk to an agent can use them — including someone shared in as a Viewer.
Listen — finish your thought first
Listen queues what you type instead of sending it. The agent stays quiet while you add more messages; when you switch Listen off, they're delivered in order and the agent responds to the whole picture.
Use it when the request needs three or four messages to explain and you don't want the agent charging off after the first one. It's the antidote to "no wait, also…".
Listen is a client-side hold: nothing reaches the agent until you release it, so queuing costs nothing.
Plan — think it through, don't do it
Plan tells the agent to analyse the request and propose an approach without carrying it out. For that turn the agent will:
- Work out what the request would actually involve.
- Lay out a step-by-step approach.
- Surface the options, trade-offs, risks, and open questions.
…and then stop and wait for you.
In Plan mode the agent won't make changes, call mutating or external-facing tools, or take irreversible actions. Use it before anything you'd rather review first — a bulk edit, a message to a customer, a repository change.
Go — carry it through
Go is the opposite instruction: proceed to completion, make reasonable assumptions, and don't check in at every step. The agent will only stop for a genuinely blocking question — a destructive or irreversible action, a security risk, or a decision it can't reasonably infer.
Use it for work you've already scoped, where the interruptions cost you more than the occasional wrong assumption.
Which to use
| Situation | Mode |
|---|---|
| The request takes several messages to explain | Listen |
| You want to review the approach before anything happens | Plan |
| The task is clear and you want it finished, not narrated | Go |
| Ordinary back-and-forth | none |
Plan then Go is a good pairing: get the plan, read it, switch to Go, say "do it."
What modes don't change
Modes apply to a single turn's behaviour. They don't change the agent's instructions, its tools, or its permissions — a Plan-mode agent has the same tools available, it's simply told not to use the mutating ones yet. And they have no effect on budgets: a Go-mode task that hits a limit stops like any other.