The question we get asked most about a personal agent isn’t what it can do. It’s a more uncomfortable one: “What if it does something I didn’t want?” It’s the right question. An agent with a terminal, files, email, and social media access can save you hours, and for the same reason it can make mistakes at the same speed. Delegating well isn’t about trusting more: it’s about deciding in advance where its autonomy ends.
This is what we’ve learned from working with our own agent every day. No jargon and no recipes: just the principles.
First, an agent isn’t an employee who goes home at night
A new employee arrives with their own judgment, common sense, and a fear of being fired. An agent arrives with capability and an eagerness to help, and nothing more. If you don’t tell it where the limits are, it invents them, and it invents them with total confidence.
That’s why control isn’t improvised in the moment of a mistake. It’s written down beforehand, in a place the agent reads every time it starts working. What isn’t written down doesn’t exist, as far as the agent is concerned.
Three categories of action, not one
The most common mistake is thinking in terms of “allowed” and “forbidden.” Real life has a middle ground, and that’s where most of the value lies. We split actions into three groups:
- What it does on its own. Reversible, internal, low-risk things: reading, organizing, preparing a draft, looking up information, taking notes. If it goes wrong, it can be undone and nobody saw it.
- What it does and tells you about. Actions with an effect but a proven routine: a recurring task that’s already been agreed on, a scheduled notice. It acts, and you know instantly.
- What it asks about first. Anything that goes out to the world or can’t be taken back: publishing, emailing a third party, spending money, deleting something, touching client data.
The practical rule for sorting an action is a single question: can it be undone without anyone noticing? If the answer is yes, let it act. If it’s no, it checks with you first.
The things it never does on its own
Beyond the prior-approval questions, there’s a short list of actions that fall outside the agent’s reach even if you seem to have asked for them in passing:
- Moving money or committing to an expense.
- Writing to third parties on your behalf unless the text or template was agreed on beforehand.
- Deleting or overwriting anything without first looking at what it is.
- Sharing private information with an external service that hasn’t been expressly authorized.
- Changing its own rules. The agent can propose an adjustment; accepting it is your decision.
That list is short on purpose. A list of fifty prohibitions won’t be read or followed; five clear rules will.
Approval should be cheap
Here’s the nuance many people miss. If every step requires a confirmation, you end up approving by reflex, without reading, and the control becomes decorative. Worse still: you get tired and turn the notifications off.
A good approval mechanism has three properties:
- It’s answered in seconds. A yes or a no from your phone, not a meeting.
- It shows what’s about to happen, not what already did. What’s useful is seeing the draft before it’s sent, not a notice that it already went out.
- It’s reserved for what matters. If it asks you a thousand times about small things, you have no attention left for the one that really counts.
Control also means knowing when to stay quiet
Control isn’t only about holding the agent back. It’s also about following your pace. An agent that offers an opinion on every message while you’re still explaining something takes the lead of the conversation away from you. In our case there’s a clear rule: when you ask it to wait, the agent answers with a single word and stays silent until you lift the hold. It sounds like a minor detail. It isn’t: it’s the difference between feeling like you’re in charge and feeling like you’re competing with your own assistant.
The same goes for the length of its answers. An agent that hands you three paragraphs when you wanted a yes or a no is imposing its pace on you. Preferences like these get written down too, and corrected every time they’re broken.
Have it leave a trail of everything
There’s no control without visibility. An agent that acts without keeping a record forces you to trust blindly, and blind trust is exactly what we’re trying to avoid. Three things should be available to check at any time:
- What it did today.
- What it decided and why.
- What’s still pending.
With our agent, that’s summarized every day in a short note. When something looks off, you open that note and understand what happened in a minute. And when a behavior is corrected, the correction is recorded with the date and the exact words, so the rule doesn’t get lost over time.
When it makes a mistake, and it will
No design eliminates mistakes; what changes is their size. An agent with well-placed limits makes small, reversible mistakes: a text that’s too long, a misread voice message, one notification too many. An agent without limits makes big ones.
Our practice when something goes wrong is always the same: say so plainly, undo it if possible, and write the rule so it doesn’t happen again. The failure isn’t hidden or justified; it’s turned into a new rule. Over time, the agent comes to look more and more like the way you want it to work, precisely because every correction is saved.
Limits move with trust
A limit isn’t a life sentence. At the beginning it’s best for the agent to ask more than necessary, because it doesn’t yet know how you think. As the weeks go by, the tasks you always approve without changing a comma move into the group it does on its own. The movement goes one way at a time and always by your decision: a rule is loosened when there’s evidence it works, not when the agent insists it’s ready.
And the other way around: if a task goes wrong twice, it goes back to the group that checks first. That flexibility is what lets you start cautiously without staying there forever.
A summary to take with you
- Write the limits before you need them.
- Separate what it does on its own, what it reports, and what it checks first.
- Use one filter question: can it be undone without anyone noticing?
- Keep the list of absolute prohibitions short.
- Make approving cost seconds.
- Require a record of what it did.
- Turn every mistake into a rule.
Delegating to an agent isn’t letting go of the wheel: it’s putting it in writing. Whoever does it well ends up with more control than before, because for the first time their rules are defined, followed, and open to review.
If you’re interested in having an agent that works for you within your limits and at your pace, get in touch and let’s talk about what yours would look like.
This article describes working principles drawn from our daily experience with our own agent. It includes no figures because we haven’t rigorously measured any result we could state in numbers.
Versión en español: Cómo darle control a un agente sin perder el tuyo
