Operate
Agent Manager
Talk to it from any screen. It knows your setup. Nothing changes without a diff you approved.
What it is
The agent behind “Talk to your agent” has a job description, not just a chat window. It lives in your sidebar as a small orb — click it from any screen and it opens already knowing your setup: where the AI runs, when it replies on its own, what changed recently. Ask it to adjust something and it does the work; ask why something happened and it answers from your real configuration, not a manual.
It never edits anything directly. Every change it wants to make arrives as a diff — what’s added, what’s removed, nothing clipped — and applies only when you approve. Every applied change is recorded, with a one-click undo. And you decide how much rope it gets: the “What your agent may operate” panel sets each capability to Off, Propose (it suggests, you click), or Auto within limits you set.
Some things it can never touch, no matter what anyone types at it: billing, team roles and data deletion aren’t settings switched off — they’re not grantable, by construction. And when you ask for something outside its reach, it doesn’t shrug “I can’t.” It walks you to the right screen with exact steps, so the conversation still ends with the thing done.
How you’d use it
Careful operator
Everything stays on Propose. The agent drafts the change, you read the diff, you click. The speed of delegation with none of the surprise — and undo behind every approval anyway.
Mid-task, any screen
You’re in Reports and notice email replies dragging CX down. Ask right there — “pause AI replies on email” — no hunting through settings. The diff appears, you approve, back to the chart.
Growing team
The capabilities you’ve watched it propose correctly for a month graduate to Auto-within-limits. The judgment calls stay on Propose. Delegation with a dial per capability — not one trust-it-or-don’t switch.
The skeptic
Tell it to cancel the subscription or delete a teammate. It can’t — not won’t, can’t: those capabilities don’t exist for it to be granted. What you get instead is the exact path to do it yourself.
Get more out of it
Start everything on Propose
Auto is earned, not assumed. Run a capability on Propose until its suggestions have been boringly correct for a while, then promote that one — the dial is per capability, so trust is granular.
Read the diff, not the vibe
The proposal card shows every line — added, removed, and what stays for context. Ten seconds of reading is the entire cost of delegation done safely.
Treat approvals as reversible
They are. Every applied change is recorded with a one-click undo, so a change that looked right at noon and wrong at three is a three-second fix, not an incident.
Ask it why, not just what
It explains its own behavior from your live configuration — why an answer went out, what widening its replies affects. The best settings screen is one you can interrogate.
Let it give directions
When it says something isn’t its to change, follow the steps it hands you — it names the exact screen and clicks. “Not my permission” never means “not your problem.”
Hire a manager, not a chatbot.
Your workspace takes about a minute to create — no sales call, no credit card.