Principal-Agent Problem

How the principal-agent problem works

The setup has three parts: a principal who owns an outcome, an agent who can take an action that affects the outcome, and a contract that pays the agent based on something observable.

1. The agent chooses an effort level e. High effort is costly to the agent; low effort is cheap.
2. Output y depends on effort plus noise: y = e + ε. Noise comes from luck, market conditions, weather — anything outside the agent's control.
3. The principal sees y but not e. The contract w(y) must therefore pay on output.

Two extreme contracts illustrate the trade-off. A fixed wage w(y) = w shields the agent from noise but provides no incentive to work hard. A full ownership contract w(y) = y − r (agent keeps everything above a fixed rent r) gives perfect incentives but forces the agent to absorb all the noise. Real contracts sit between: a base salary plus a performance fraction.

The optimal incentive intensity falls when noise is high (agent is bearing noise unnecessarily) and rises when the agent's effort is highly responsive to incentive (so a smaller bonus moves more effort). This is the heart of the Holmström-Milgrom (1987) linear-contract characterization.

When the framework applies

• Delegation with effort that's hard to observe. Sales, surgery, software engineering, parenting.
• Outcomes are noisy proxies for effort. Quarterly profit, exam scores, health outcomes.
• The principal cares about the outcome. Otherwise no contract problem.
• The agent is risk-averse. Without risk aversion, full ownership is always optimal.
• Long-run, repeated, or otherwise enforceable contracts. One-shot anonymous deals collapse under moral hazard.

Compensation schemes compared

	Flat salary	Piece rate	Bonus & threshold	Stock options	Restricted stock with vesting	Tournament (promotion)	Sharecropping
Effort incentive	None	Strong, linear	Strong, kinked	Strong, convex	Strong, concave	Strong, ordinal	Linear, partial
Risk on agent	None	Heavy	Heavy near threshold	Heavy at low strike	Less than options	Relative — filters macro noise	Half of weather risk
Encourages risk-taking	No	Neutral	Yes near threshold	Strongly yes	Mildly yes	Yes — gambling for promotion	Neutral
Distortion if multitask	Low effort overall	Game the metric	Sandbag below, push above	Excessive risk	Modest distortion	Sabotage rivals	Underinvest in soil
Long-run alignment	None	None	Quarterly only	Until expiry	Through vesting	Career-long	Until lease ends
Where used	Civil service, junior staff	Garment piecework, Uber	Sales bonuses, executive STIs	Tech CEOs, Silicon Valley	Modern executive comp	Law-firm partner track	US South historically
Theoretical reference	Holmström 1979	Lazear 1986	Healy 1985	Yermack 1995	Hall-Murphy 2003	Lazear-Rosen 1981	Cheung 1969

No scheme dominates. The optimal mix depends on how noisy output is, how risk-averse the agent is, how many tasks the agent juggles, and whether the principal can observe relative performance.

Worked example: stock options vs salary

Suppose a CEO can choose effort e ∈ {0, 1}. Effort cost: c(e) = $2M if e = 1, $0 if e = 0. Firm value: y = e × $1B + ε, where ε is normally distributed with standard deviation $400M (about 40% of mean — realistic equity volatility).

The CEO is risk-averse with constant absolute risk aversion (CARA) coefficient r = 4 × 10⁻⁹ per dollar. The board considers two contracts.

1. Flat salary: $10M regardless of y. CEO chooses e = 0 (no incentive to exert costly effort). Firm value: $0 + ε. Expected board profit: 0 − $10M = −$10M.
2. Linear option-equivalent: w = $4M + 0.01 × y. The CEO bears 1% of firm noise: variance = 0.0001 × ($400M)² = $1.6 × 10¹⁵. Risk premium ≈ ½ × r × variance = ½ × 4 × 10⁻⁹ × 1.6 × 10¹⁵ = $3.2M. CEO's expected utility from effort 1: $4M + 0.01 × $1B − $2M − $3.2M = $8.8M. From effort 0: $4M + 0 − $3.2M = $0.8M. Effort 1 dominates. Firm value: $1B + ε. Board profit: $1B − $14M ≈ $986M.

The 1% slope is enough because the effort productivity ($1B) is huge relative to its cost ($2M). With less responsive effort or noisier output, the optimal slope rises (and the risk premium rises with it). This is the second-best world: the principal pays $3.2M of dead-weight risk premium to induce $1B of effort. Nothing better is achievable when effort is hidden.

Variants and extensions

• Multi-task contracts (Holmström-Milgrom 1991). When agents juggle measured and unmeasured tasks, strong incentives on the measured task distort effort. Optimal contract may be deliberately weak.
• Career concerns (Holmström 1982/1999). Even without explicit incentives, agents work hard to build a reputation that future employers will reward. Implicit incentives substitute for explicit ones early in a career, then weaken.
• Tournaments (Lazear-Rosen 1981). Promote whoever produces most relative to peers. Filters out shared noise; introduces sabotage and risk-taking.
• Subjective evaluation (MacLeod 2003). When output is hard to verify, principals use discretionary bonuses backed by reputation. Common in white-collar work.
• Relational contracts (Levin 2003). Repeated interaction supports incentive contracts even without enforceable formal terms.
• Hierarchies and supervision (Tirole 1986). Add a third party — a monitor — and you get coalitions, collusion, and a richer set of incentive problems.
• Public-sector agency. Voters as principals, politicians as agents; multiple principals, multiple tasks, infinite horizons. Most public-policy reforms are principal-agent fixes (oversight boards, term limits, sunset clauses).

A brief history

The terminology dates to Stephen Ross's 1973 AER paper "The Economic Theory of Agency: The Principal's Problem" and Michael Jensen and William Meckling's 1976 JFE paper "Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure." The mathematical foundations came from James Mirrlees, who developed the technical machinery for hidden-action contracting in unpublished 1971 notes and a 1976 paper.

Bengt Holmström's 1979 Bell Journal paper "Moral Hazard and Observability" stated the informativeness principle: any signal correlated with effort but not redundant given output should enter the contract. Sanford Grossman and Oliver Hart's 1983 Econometrica paper "An Analysis of the Principal-Agent Problem" formalized the general two-period model with discrete effort and outputs.

Mirrlees won the 1996 Nobel Prize, sharing with William Vickrey for foundational work on incentives under asymmetric information. Holmström and Oliver Hart shared the 2016 Nobel Prize "for their contributions to contract theory" — the citation explicitly named the principal-agent framework. Jean Tirole won in 2014 for industrial organization that drew heavily on agency theory. The framework is arguably the dominant analytical lens of modern microeconomics.

Common pitfalls

• Designing contracts on raw output instead of relative performance. Violates Holmström's informativeness principle whenever peer comparisons are available — and they almost always are.
• Ignoring multitasking. Strong incentives on what you can measure starve what you cannot. Teachers teach to tests; surgeons cherry-pick easy patients.
• Treating risk-averse agents as risk-neutral. Loading too much variance onto an agent extracts a risk premium far in excess of the incentive effect. CEO option grants are notoriously over-volatile.
• Forgetting the participation constraint. The agent must be willing to take the contract — pay a base sufficient to clear their outside option, or no contract is signed.
• Assuming long horizons fix everything. Repeated interaction helps via career concerns but introduces ratchet effects (today's high performance raises tomorrow's quota), which themselves distort effort.
• Not auditing the audit. Adding a supervisor creates a second principal-agent layer; the supervisor needs incentives too. Tirole's 1986 collusion problem.
• Confusing alignment with control. Aligned incentives don't replace monitoring entirely; both are partial substitutes, with the optimal mix depending on cost.