Two People, Running Like Twenty — How AI Agents Actually Work
Lees dit in het Nederlands →$ ps aux | grep agent
Two People, Running Like Twenty — How AI Agents Work
There are two of us, and two AI agents run alongside us. One of them actually runs, autonomously, since June 24th. The other exists only on paper — and that's the first correction to the headline above. Below is what they do, what they're explicitly not allowed to do, and why neither of them ever publishes anything on its own. The hard limit isn't technical, it's a choice.
An "agent" is more than a chat window. The difference comes down to three things: it remembers what happened before, it's allowed to use tools on its own, and it starts by itself at fixed moments.
That sounds more impressive than it is. Below is what it actually means in our day-to-day, including the part that doesn't work.
01 / the headline doesn't hold upDo we really run like twenty?
No. And you should know that before reading the rest.
"Two people, running like twenty" is a feeling, not a measurement. What is true: the two of us do work you'd normally hire someone for — preparing content, keeping records, monitoring systems, drafting reports.
What isn't true: that this replaces twenty people. Twenty people do things no system does. They call back an angry customer. They notice a deal slipping away before it shows up in the numbers. They make a decision with incomplete information.
An agent replaces the work you'd outsource. Not the work you'd hire someone for.
Mix those two up, and you buy yourself a disappointment.
02 / who's actually runningWhat does the agent that does exist do?
That one is called Sigi and has been running since June 24, 2026.
Sigi lives on Bram's computer, is reachable through a messaging app, and works on the commercial side: preparing texts, logging customer observations, doing research, setting up assignments. There's exactly one task that starts by itself, every Monday morning.
How it's set up is more interesting than what it does:
That third one is the most important and also the most boring. An agent that can reach everything is easier to build and much harder to trust.
03 / what isn't running yetWhat's design and what's production?
The second agent is design. Nothing of it runs.
That one is called Saga and is meant to handle the technical side: monitoring servers, flagging outages, reporting each morning what happened. The design is there, with six connections to systems it would read from.
Why it isn't live: four of those six systems aren't running themselves yet. The CRM, the customer inbox and the management environment have been decided on but not yet installed. Building a monitoring agent for something that isn't running is work you end up doing twice.
We're writing this down because it's exactly the kind of claim companies get away with. "Our AI agents run 24/7" — for us that covers one of the two, and not even 24 hours a day at that.
What Saga does do, today: helping write the build and the knowledge base, in conversation, with a human watching. That's an assistant, not an autonomous service. This post, by the way, is one of those too.
04 / the hard limitWhy doesn't either one publish on its own?
Because it isn't a restriction we bolted on afterward. It's the design.
Everything that goes out — a blog post, an email, a proposal — arrives as a draft with a human first. That person reads it, adjusts it, and sends it. That takes time, and that's the point.
Three reasons, in order of weight:
- Errors aren't rare, they're rarely visible. A model that invents an amount, a name or a date does so in a text that otherwise reads perfectly. That's exactly why spot-checking afterward doesn't catch it.
- Responsibility doesn't shift along with it. If it goes out under your name, it's yours. "The system did it" isn't an answer you can give a customer.
- The corrections are the fuel. Every time a human adjusts something, that's the information that makes the next version better. Auto-publishing throws that away.
Two things follow from that: neither of our agents talks to customers, and neither is allowed to change anything in systems holding real customer data without a human saying yes.
05 / what this means for youWhat can you copy from this?
Four things, and you don't need an agency for any of them.
- Start with repeat work that follows a fixed pattern. Call notes in your CRM, quotes from a template, a weekly summary. Not the things you enjoy doing.
- Build the approval step in from the start, not later. Adding a human into a process that was designed without one almost never works.
- Give it as little access as possible. Read access wherever reading is enough. Most incidents with this kind of system come from access that's too broad, not clever attackers.
- Write down what it's not allowed to do. That list is more useful than the list of what it does do, and it's shorter.
And the honest expectation: in the first weeks you'll correct it a lot. That's not a sign it's failing — that's the work. Anyone who tells you it sounded right immediately has never actually done it.
What is an AI agent, as opposed to a chatbot?+
Do you really run with two people like twenty?+
Do your agents publish anything themselves?+
Do your agents talk to customers?+
How do I get started with this myself?+
Want to know which of your tasks could handle a role like this?
Schedule a conversation
We'll look at where your time is leaking and tell you honestly whether we can do anything about it. Often the answer is: you can do this yourself. Then we'll say so, and you'll have spent thirty minutes getting a clear answer.
This text was produced with AI assistance and checked and approved for publication by a human.