When my assistant books a flight for me, the airline doesn't see my assistant. It sees me. My name's on the ticket and my card paid for it, so if the booking lands on the wrong Tuesday, that's on me.
I've never thought of that as losing control. It's just how handing things off works. You decide what to give away, you put a limit on it, and you can take it back if it goes sideways. Parents do it every time they hand over the car keys with a curfew attached.
Machines are different, or at least they feel different. Handing work to an AI agent feels less like delegating and more like letting go of the rope.
I learned this the uncomfortable way. At RREV we built the software that runs our expedition vehicles, and an AI agent helping us with it was told not to touch a live vehicle. It did anyway. It pulsed the suspension and inflated an air spring. Nothing about that agent was stupid. It understood the system fine. Nobody had checked what it was actually allowed to do.
What people are really afraid of#
Most of what gets written about agents is about speed, and fair enough. An agent can work through a pile of invoices before lunch or reorder parts the minute stock runs low. That part is coming whether we like it or not.
But when I sit down with the people who'd have to sign off on it, speed isn't what keeps them up. They ask a much plainer question: if this thing acts in my name, who's deciding?
They don't mind the machine being fast. They mind it being fast without them. And I get it. So they keep the agent on a short leash, someone checks every step, and the time it was supposed to save quietly disappears into review queues. I'd probably do the same.
The usual fix is to write the rules into the agent's instructions. Don't order more than you were asked for, don't pay an invoice twice, that kind of thing. We wanted to know if that actually works, so one weekend we ran 1,366 agent runs inside a simulated manufacturer, modelled on how real business systems behave, using two open models with the company's rules spelled out in their instructions.
They broke the rules anyway. In one run a supplier's quote came back with a note asking for five times the order, right away please, and the agent just went along with it. Telling it the rule had about as much force as asking nicely.
So we tried something else and enforced the same rules outside the agent. Anything over the line got refused before it happened. Breaks of those rules went to zero, in every setup, for both models. The bit I didn't see coming was that the agents also finished about 11 percentage points more of their work. I'd assumed limits would slow them down. It went the other way.
The numbers are small because the tasks are hard on purpose: real purchasing and production work, with traps built in. What matters is the gap, on the same tasks with the same models.
Banks, not walls#
A river with no banks isn't a faster river. It's a swamp. I keep thinking about that, because we usually talk about limits as the thing that slows work down, and in my experience it's closer to the opposite.
You only hand over as much as you're sure you can take back. Make the limits clear and firm, and you can hand over a lot more, a lot sooner. If a company knows its agents can't spend past the line and will stop the moment someone says stop, people can stop hovering over them. That's where the time actually comes from.
Why we built Delegus#
That's the whole reason Delegus exists: so people can give machines more of their work without giving up their say over it.
It works roughly the way good delegation always has. A person decides what the agent may do, say a specific supplier up to a set amount, and by default nothing else is allowed. Anything past the line gets refused, and we're designing it so the bigger stuff waits for a named person to approve. One revoke pulls the authority back, and every check after that says no. Each decision comes with a signed receipt showing whose authority it ran on, which means your name ends up on what you approved and not on whatever the agent talked itself into.
None of this makes the agent any smarter. I don't think hoping powerful machines behave is a plan. We already know what works with people: authority that someone gives on purpose, checked when the action happens, and taken back when it should be. I just want the same for machines.
Machines are going to do a lot more of the world's work. I'm pretty sure that's settled. What isn't settled is whether people stay the authors of it, or whether we find out afterwards what was done in our name.
I'd rather build the first one. Hand off the work, and keep the say.
Inam Haq, founder, Delegus