Agents that ask first

The pitch for an AI agent is autonomy: give it a goal, and let it work while you do something else — or nothing at all. The fear is the same sentence read from the other side. An agent that can act on your behalf, with your sign-ins, while you sleep, can also make mistakes on your behalf, with your sign-ins, while you sleep.

Both are true, and the difference between a product you can use and a demo you can’t is a single design decision: which actions need a human yes.

The line is irreversibility

Not all actions are equal, and treating them as if they were is where agent products go wrong — either by asking about everything (and becoming a very elaborate notification machine) or by asking about nothing (and becoming a liability).

The line we draw is simple. Work that stays inside the machine is free: reading, searching, drafting, organising, preparing. Work that crosses the boundary — that changes something in the world or in someone else’s system — asks first:

  • Sending an email or message
  • Posting anything public
  • Deleting files or data
  • Paying for anything
  • Changing an account or its settings

An approval is not a vague “the agent wants to do something.” It shows the exact email, the exact post, the exact amount — the thing itself, ready to go — and it goes nowhere until you say yes. If you say no, the agent adjusts or stops. Nothing about the step is hidden in a log you were supposed to check later.

Approvals are what make always-on possible

It is tempting to see approval as a brake on autonomy. It is the opposite. The reason you can let the Attovex teammate work overnight — the reason Attovex machines can be persistent, signed in, and left running — is that the dangerous verbs are gated. You are not trusting the agent to be perfect. You are trusting a system in which imperfection is survivable, because nothing irreversible happens without you.

That is also how trust grows in the right direction. You start by approving everything and reading closely. Over time you learn which routines are reliably right, and the relationship becomes what it is with a good human assistant: broad freedom on the work, a check-in before anything that can’t be undone.

Asking first is a feature, not a disclaimer

Many products bury this in terms and settings. We put it in the first sentence of what Attovex is: an always-on AI teammate for your business that runs on your Mac and, when your Mac is off, on an Attovex machine — and always asks before it sends, posts, deletes or pays. It is in our privacy policy and our terms, not as legal cover but as the product working as described.

Autonomy you can’t leave alone isn’t autonomy; it’s supervision with extra steps. An agent that asks first is one you can actually give the night shift to.

Attovex is in early access. To ask about joining, email [email protected].

Frequently asked questions

What does 'asks first' mean in practice?

Before an Attovex teammate takes an action that leaves the machine — sending an email, publishing a post, deleting a file, making a payment, changing an account — it shows you exactly what it is about to do and waits for your approval. Reading, drafting and researching happen freely; consequences need a yes.

Won't approvals slow the agent down?

Approvals apply to a small minority of steps: the irreversible, outward-facing ones. The agent still works autonomously on everything else — gathering, drafting, organising — and queues approvals so you can clear them in one pass, including from your phone. What you approve is the exact message or change, not a vague summary.

Why not just let the agent act and review afterwards?

Because some actions can't be reviewed back into existence. A sent email, a published post and a completed payment are facts in other people's systems. Undoing them ranges from embarrassing to impossible, so the check has to happen before, not after.

← All posts on the Attovex blog