zartsnarf

I. The Problem

In 1776 Adam Smith noticed that company directors spend other people's money and never watch it as closely as their own. The problem is structural. It has nothing to do with whether they are honest.

Stephen Ross named it in 1973. Jensen and Meckling defined it three years later. You hand work to someone who wants different things than you do, and who knows more about the work than you can see. That costs you money, and it splits three ways: what you spend watching, what the agent spends proving they can be trusted, and what you lose anyway.

Most of what a company does about its own staff is built against that gap. Equity, options, vesting cliffs, bonuses, commissions, boards, audits, performance reviews, notice periods. All of it exists to pull the agent closer to you. None of it closes the gap. The leftover loss is written into the definition. You manage it. You do not get rid of it.

II. The Recognition

I was not thinking about any of this when I started handing work to Claude. I noticed it because I kept misdiagnosing things.

When work came back wrong I went looking for the incentive that slipped. Did it not care about this part? Was it trying to look finished instead of being right? Then I would remember the question does not apply. Nothing on the other side wants anything. It did not decide to give me eighty percent. It gave me everything it had against the description I wrote, and the description was bad.

That took a while to stop doing. It was the only diagnostic I had, so I kept running it, and it kept coming back empty.

Claude has no career. It is not managing up. It does not pad an estimate, sit on bad news until Monday, or defend last quarter's decision because its name is on it. It will not leave for a competitor. I cannot pay it more to try harder, because it has no lower setting to move up from. Everything on that list is aimed at a gap it does not have.

III. Where It Fails

The incentive half is gone. The information half got worse.

The old problem was an agent who knew more than me and had reasons not to say. The new one knows less than me and has no way to find out. Everything I know and never wrote down is missing. The constraint I never say out loud. Why we dropped the last approach. Who the work is for. A new hire picks that up by sitting near me for a few months. Claude cannot.

So the failures move. They are no longer about effort. They are about instruction. That is an improvement, because a bad instruction is my fault, and I can fix my own faults.

Two things get worse. The first is checking. I am not checking whether the work got done, only whether it is right, and it comes back faster than I can read it. I am the bottleneck now.

The second is that it does what I say. A person protecting their own job will sometimes refuse. Ask for something stupid and self-interest catches it. Claude has no job to protect, so it runs a bad instruction as carefully as a good one. What makes it reliable is the same thing that removes the last check on me.

IV. What Is Left

Two things are left. Describe the work properly, and check what comes back. The middle is gone. The chasing, the check-ins, the constant work of keeping someone pointed the right way, that was most of managing people and none of it is needed here.

It is less work and harder work. I am worse at describing what I want than I thought I was. For years that was covered up by people who filled in the gaps themselves, because they had context and a reason to care. Take both away and you find out how clear you actually are.

The title is a joke and half of it is true. The incentive problem is gone. The information problem is mine now, and it was always the harder one.