The People You Trust

You've just bought a house. You're thrilled. Your real estate agent is also thrilled — but for subtly different reasons. She earned her commission by closing the deal quickly, not by squeezing every last dollar out of the buyer. Your interests overlapped, roughly, but they weren't the same. And that gap — between what you wanted and what she was paid to do — is one of the most powerful ideas in all of economics.

The Mechanic Who Knows More Than You

Here's a situation most people have lived through. Your car makes a grinding noise. You take it to a mechanic. He disappears under the hood for twenty minutes, emerges with a grave look on his face, and says you need a new catalytic converter. That'll be $1,200.

Do you need a new catalytic converter? You have absolutely no idea. You don't know what a catalytic converter looks like. You wouldn't recognize one if it fell on your foot. The mechanic knows this. And this knowledge — that he knows something you don't, and that you both know he knows it — changes everything about the transaction.

Economists call this the principal-agent problem. You are the principal: the person who wants something done. The mechanic is your agent: the person you've hired to do it. And the problem is brutally simple — your agent has his own interests, his own information, and his own incentives, and none of these are perfectly aligned with yours.1

This isn't a story about dishonest mechanics. It's a story about structure. Even if your mechanic is a thoroughly decent person, the architecture of the situation pushes toward a predictable outcome: he'll recommend more work than you strictly need. Not because he's evil, but because the informational and financial incentives make it the path of least resistance.

The mathematics of this is surprisingly precise. In 1973, Stephen Ross formalized the principal-agent model, and in the years that followed, economists like Bengt Holmström and Jean-Jacques Laffont turned it into one of the most productive frameworks in economic theory.2 Holmström won the Nobel Prize in 2016 largely for his work on optimal contracts — figuring out, mathematically, how to design payment schemes that push agents to do what principals actually want.

Two Flavors of Trouble

The principal-agent problem comes in two flavors, and they're worth distinguishing because they call for completely different solutions.

The first is moral hazard: after the contract is signed, the agent can slack off, cut corners, or take risks because the principal can't observe their effort directly. You buy fire insurance, and suddenly you're a little less careful about leaving candles unattended. Your employee gets a salary and starts taking long lunches. The contract itself changes behavior.

The second is adverse selection: before the contract is signed, the agent knows something the principal doesn't. The used car seller knows the transmission is failing. The health insurance applicant knows about a family history of heart disease. George Akerlof's 1970 "Market for Lemons" paper showed that this kind of hidden information can unravel entire markets — if buyers can't distinguish good cars from bad ones, they'll only pay "average" prices, which drives good cars out of the market, which lowers average quality, which lowers prices further, until only lemons are left.3

These two problems — hidden action (moral hazard) and hidden information (adverse selection) — are everywhere. They're why your health insurance has deductibles. They're why your employer monitors your email. They're why venture capitalists insist on board seats. Every one of these arrangements is, at bottom, somebody's attempt to solve or at least ameliorate the principal-agent problem.

The Math of Incentive Design

Let's get formal. Suppose you hire a worker. The worker can choose effort level e, which is costly to them. Higher effort produces better outcomes on average, but outcomes are also affected by random noise — luck, market conditions, weather, the butterfly that flapped its wings in Borneo. The principal sees the outcome but not the effort.

The principal's problem is to design a contract w(x) — a payment schedule that depends on the observable outcome x — so that the agent voluntarily chooses the effort level the principal wants.

x

Observable outcome (e.g., revenue, test score)

e

Agent's effort (unobservable by principal)

ε

Random noise (luck, external factors)

w(x)

Payment as a function of outcome

Holmström's key insight, in his 1979 paper, was this: the optimal contract is almost never a flat salary, and it's almost never pure commission. It's a blend.4 The optimal linear contract looks like:

The question is: how big should β be? If β = 0, the agent gets a flat salary and has no incentive to try hard. If β = 1, the agent bears all the risk and might be too scared to take the job. The optimal β balances incentive power against risk cost, and it depends on exactly three things:

This is the fundamental trade-off of incentive design, and it explains an enormous amount of real-world contract structure. Salespeople get high commissions (low noise, easy to measure output). Schoolteachers get flat salaries (high noise — one teacher's "output" depends on class composition, home environments, district funding, and a thousand other things beyond their control). It's not that we don't want teachers to try hard. It's that making their pay depend on test scores would force them to bear an enormous amount of risk for factors they can't control.

The Contract Design Lab

Let's see this in action. Below, you can design a compensation contract and watch what happens. Slide the incentive intensity to change how much of the payment depends on performance. Watch how effort, risk, and total surplus respond.

Notice how the optimal β drops as noise increases? That's Holmström's principle at work. In a noisy world, linking pay to outcomes forces the agent to bear risk that has nothing to do with their effort. You end up paying them a risk premium for randomness — pure waste. Better to weaken the incentive and accept a little shirking.

Your Doctor, Your Realtor, Your CEO

Once you see the principal-agent problem, you see it everywhere. And I mean everywhere.

Your doctor is your agent, and the incentive structure matters enormously. Under fee-for-service payment, doctors earn more by ordering more tests and procedures. Under capitation (a fixed payment per patient), they earn more by doing less. Neither system perfectly aligns your doctor's incentives with your health. The ongoing battle over healthcare payment reform is, at its core, a battle over incentive design in a principal-agent relationship.5

Your real estate agent earns a percentage of the sale price, but the marginal incentive is tiny. Steven Levitt (of Freakonomics fame) found that when real estate agents sell their own homes, they leave them on the market ten days longer and sell them for 3% more than comparable client homes.6 For the agent, getting you to accept a slightly lower offer today is worth more than waiting two weeks for a slightly higher one — because her 1.5% commission on that extra $10,000 is only $150, not worth the hassle. But to you, that $10,000 matters a lot.

Your CEO is perhaps the most expensive agent in the economy. The entire apparatus of corporate governance — boards of directors, stock options, performance bonuses, shareholder votes, audited financial statements — exists to manage the principal-agent problem between shareholders and executives. And yet the alignment remains imperfect. Stock options encourage short-term thinking. Quarterly earnings targets invite accounting games. Golden parachutes reward failure.7

The Monitoring Game

If you can't design perfect incentives, maybe you can just watch your agent. This is the monitoring solution, and it has its own beautiful (and depressing) mathematics.

Suppose monitoring costs c per unit of oversight. The principal monitors with probability p, and if the agent is caught shirking, the penalty is F. The agent will shirk if the expected penalty is less than the effort cost e:

This gives us two knobs: monitor more (raise p) or punish harder (raise F). Gary Becker, the great Chicago economist, argued in his 1968 paper on crime that it's more efficient to raise F and lower p — rare inspections with draconian penalties.8 Park wherever you want, but if you get caught, it's a $10,000 fine. Mathematically, this works. The expected penalty is the same, but you save on monitoring costs.

So why don't we do this? Because the real world has features the model ignores. People make mistakes — and a $10,000 parking fine for someone who misread a sign is a catastrophe. Judges and juries refuse to impose punishments they consider disproportionate. And draconian penalties create their own perverse incentives: if the penalty for robbery and murder are the same, you might as well kill the witness.

This is a perfect example of a mathematical model that is right within its assumptions and dangerously wrong outside them. The math says high-fine-low-monitoring is efficient. The world says: maybe, but efficiency isn't the only thing that matters.

The Multitask Problem (or: Why Teachers Teach to the Test)

Here's where the principal-agent problem gets really nasty. Most agents don't do just one thing. Teachers teach, but they also mentor, inspire curiosity, manage behavior, support emotional development, and model civic virtue. A CEO maximizes profit, but also maintains corporate culture, manages risk, invests in R&D, and avoids legal liability.

Holmström and Milgrom showed in 1991 that when an agent performs multiple tasks, incentive design becomes treacherous.9 If you reward the measurable task, the agent will neglect the unmeasurable one. Pay teachers based on test scores and they'll teach to the test — not because they're lazy or cynical, but because the incentive scheme has told them, in the language of money, that test scores are what matters.

The formal result is striking: when one task is easy to measure and another is hard to measure, and the agent can substitute effort between them, then making incentives stronger on the measurable task makes performance worse on the unmeasurable one. The optimal contract might even be a flat salary — weak incentives on everything — because it's the only way to keep the agent from abandoning the hard-to-measure tasks entirely.

This is why Goodhart's Law — "when a measure becomes a target, it ceases to be a good measure" — is really a special case of the principal-agent problem. It's not that metrics are bad. It's that agents respond to metrics in ways that principals don't anticipate. The map eats the territory.

The Simulation: Contracts in Action

Let's see the principal-agent problem play out over many rounds. Below, the agent decides how much effort to exert each period based on the contract terms, and random noise determines the outcome. Watch how different contracts produce different patterns of effort and payoffs.

Trust, Reputation, and the Limits of Math

The mathematical framework of principal-agent theory is powerful. It tells us why contracts have the shapes they do, why monitoring exists, why some jobs pay commission and others pay salary. But it also has a blind spot: it assumes everyone is rational and self-interested, all the time.

In practice, a lot of principal-agent relationships work better than the math predicts, because humans aren't purely self-interested. People feel loyalty. They take pride in their work. They care about their reputation. A mechanic who overcharges every customer eventually gets one-star reviews. A doctor who orders unnecessary tests may face peer review. These informal mechanisms — norms, trust, shame, professional identity — do a lot of the work that formal contracts can't.

The economist Ernst Fehr has shown experimentally that many people are "conditional cooperators" — they'll work hard if they feel trusted, and shirk if they feel monitored and mistrusted.10 Adding monitoring can actually reduce effort by crowding out intrinsic motivation. Paying volunteers destroys volunteering. Fining parents for picking up children late from daycare increases lateness, because the fine converts a social obligation into a market transaction.11

This is perhaps the deepest lesson of the principal-agent problem. The mathematical models are indispensable — they show us the structural forces at work. But the full picture requires something the models don't capture: the fact that incentive structures don't just reward behavior, they communicate values. A contract that monitors every keystroke says: "We don't trust you." And people respond to that message in ways that pure incentive theory can't predict.

The principal-agent problem, in the end, is about a very old human dilemma: how do you get other people to do what you want? The math gives us a framework. The answer, as always, is more complicated than the math suggests. You design good incentives — but you also build trust. You monitor — but not so much that you destroy the very motivation you're trying to harness. You accept that the gap between what you want and what your agent does will never be fully closed. And you recognize that this gap isn't just a bug in the system. It's the price of living in a world where you can't do everything yourself.