An agent that doesn't write things down learns nothing

Every turn of an agent starts with an empty head. What it did yesterday, what it weighed up, why it left something alone: all gone, unless it is written down somewhere.

That is the whole case for a logbook, and it is less obvious than it sounds. Most people think of a logbook as control: something to look in when things go wrong. That is part of it. But an agent's logbook is mostly its memory, and with that the precondition for nearly everything you want to use agents for: handing work over, reviewing mistakes, and getting better.

In this piece I walk through what belongs in it, why you write it down in the moment and not afterwards, how a mistake becomes a lesson, and how that comes together in one place in my own setup. The examples come from my book AI-agenten die het werk doen, from Het autonome AI-bedrijf and from working with my own agents.

Not on disk, not happened

In AI-agenten die het werk doen I describe how, when I looked back for the book, I still had everything. Every change with a timestamp. Every task with the reason next to it. More than twelve hundred pieces of work where I could read why they existed.

And for the first four days I had nothing. Not because anything was lost. The changes are there to the minute. What is missing is the why: not a single consideration, not a single doubt, not a single choice with a reason attached. In those days we built and did not discuss, so nobody wrote anything down. They were the four days in which everything came into being.

The odd thing is that I knew it all at the time. Every choice had a reason, and that reason was so obvious that writing it down felt like a waste. You don't write down why you turn left when you know the way. Two weeks later the way is gone. Not the route, that is still there, but the landscape around it: what you considered, what you rejected, why this had to come first and that later.

This is the most underrated kind of loss there is, because you don't notice it when it happens. Nothing breaks. You just look at your own work a month later and think: why did I do it like that?

For an agent it is no different, only far sharper. A human still has a vague feeling about their own work. An agent has nothing, apart from what is written down.

What goes in

Everything an agent does, it writes down, in one place and in plain language. Four things belong in it:

• what it saw: the question, the source, the message;
• what it decided, with the reason;
• what it did, and what happened next;
• what it did not do, and why not.

The last one is the most useful. An agent that lets an opportunity go because it is unsure whether it fits the rules must write that down. Otherwise nobody can tell "not found" from "deliberately not done". In Het autonome AI-bedrijf this even has a part of its own, logging what did not happen, because a log that only shows what was done is missing half the story.

And the reason is worth more than the decision. Without a reason you can only follow a decision or break it. You cannot revisit it. Say it was agreed that someone comes on fixed days. If only that is written down, then in six months it is a rule nobody knows the origin of. If the why is written down too, you can see whether that reason still holds. In one case you have a habit, in the other a decision that can move with you.

In the moment, not afterwards

The obvious fix is to do it at the end: at the end of the day, the week, the project. That doesn't work, for two reasons that have nothing to do with discipline.

By the end, half of the why is already gone. You write down what you did, and that is usually recorded somewhere already. What you no longer write down is what you almost did and why you didn't, and that is exactly what you won't find anywhere later.

There often is no end. A day runs out without a wrap-up. A project changes direction. A conversation gets interrupted. And without an end there is no moment for it to happen after all.

So you write it down while it happens. Not neatly, and not in sentences anyone else would admire. Rough and present is worth infinitely more than polished and absent. A reason you write down at the end of the week is a reconstruction, and it always sounds wiser than it was.

Four uses, and you need all four

A good logbook is used in four ways:

1. Handover. Another agent or a human sees where the work stands without asking. It is the only way anything can carry on without you.
2. Review. The reviewer looks not only at the result but at how it came about. A review like that only works if the agent left traces in a form that can be read.
3. Looking back. Over a period you look at what happened often and what kept going wrong in the same way. This is not a hunt for culprits, but a search for patterns.
4. Learning. What the look back yields goes back into the rules of the role. Then you don't just have an agent that works, but one that gets better.

The fourth use is where the logbook goes from memory to growth. Without a log an agent learns nothing, however good the model, because every turn starts from scratch. With a log, a pattern that went wrong five times can become one line in the role card, and then it doesn't go wrong a sixth time.

Note a mistake the moment you find it

There is a habit that goes with this, which I apply everywhere, to myself and to my agents. If, while working on A, you find a mistake in B, note it at that moment, before you carry on. Even if you are not going to fix it.

Not at the end of the session, not "if it turns out to be relevant". A mistake you only mention in a conversation is gone once the conversation is over. A mistake that sits in the log of other work cannot be found once that work is done. And a session that stops abruptly never wraps up, so "I'll note it later" in practice means it disappears.

It doesn't have to be much: what you saw, where, and what you expected. One line at the right moment is enough, and it doesn't interrupt your work. It is a note, not a context switch.

For an agent this matters even more than for a human. It won't think halfway through a turn, "I still have to do that later". It writes it down, or it is gone.

One line: where was I

In the book I also dwell on the smallest part of the logbook, which is the most used. At the top of the work there is one line: what you are doing and what comes next. Whoever picks up tomorrow, you, a colleague or another agent, doesn't have to reconstruct where the story was left.

Two conditions make that line useful. Update it the moment you are somewhere else, not at the end, because by then half of the why is gone. And remove it when the work is done: a line that points to finished work is worse than no line, because the next person starts on something that no longer exists.

Keep it a pointer, not a report. "Working on the quote for X, waiting for Y's prices" is enough. Whoever puts more in won't keep it up.

And this isn't about software

It is not only about agents. An advisor who has everything in their head can never take on two assignments at once, can't bring anyone up to speed and can't fall ill. They call it experience. It is also a design choice, and they probably never made it consciously.

A school rediscovers every year where pupils stumble, because it was never written down anywhere. The problem is not that nobody knows it, but that the knowing lives nowhere.

Where it all comes together for me

What I described above runs in one place for me: my own Wantzel Studio. Agents, tickets, books, mail and the compiler it is built with live in one environment. The studio is my own workplace.

What does that look like in practice?

Every agent has a role, on a card. It says what the goal is, what the result should be, who checks the work and when the agent brings me in. An agent that manages LinkedIn does something different from an agent that answers my mail, and each works in its own queue with its own logbook.

The work runs through tickets. A ticket says what has to happen. What the agent does next is recorded in the ticket: what it read, what it made, what it found. If I want to know in the evening why something went the way it did, I read it back instead of guessing. The same goes for a fellow agent taking over: everything it needs is already there.

Nobody approves their own work. What an agent produces goes through a check before it goes anywhere. That applies to code, to texts and to messages to people.

Decisions about money, publishing and contact with people stay with me. An agent puts forward a proposal, I say yes or no, and only then does it go out the door. This article is an example itself: it was ready as a draft first and only went live after I had read it.

Every role has a budget, a ceiling and a list of what it does not do. That keeps the work calm. I don't have to fear that something happens overnight that I didn't know about.

Why do I build this myself? Because I want to be able to rely on what my agents do. That works best when I can see the whole chain, from the lowest layer to the finished work. If one link belongs to someone else and I can't look inside, I don't know where to look when something goes wrong. You can find more about the studio at wantzel.com. The ways of working behind it are described in AI-agenten die het werk doen: how you give agents a role, how you let them work together and where you, as a human, stay at fixed points. Every chapter has an exercise you can do the same evening.

What you can do on Monday

You don't need a studio to start. Take one task that an agent, an automation or a colleague does for you.

1. Is the why written down anywhere? Pick the latest decision in that work and look for the reason. If only the outcome is there, you know where to start.
2. Put one line above it. What is going on here and what comes next? Update it the moment you are somewhere else.
3. Note the next mistake straight away. What you saw, where, what you expected. One line.
4. Look back in a month. What happened often? Turn that into one line in the way of working.

Where is it recorded in your organisation why a decision in a process was made the way it was, and can someone else read that without calling you?

All pieces