Multitask Principal–Agent Model

The multitask principal–agent model is a foundational framework in contract theory, organizational economics, and business and economics theory introduced by Bengt Holmström and Paul Milgrom in their 1991 Journal of Law, Economics, & Organization article, "Multitask Principal–Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design." Extending the classical single-task principal–agent problem and Holmström's informativeness principle to environments where an agent allocates multidimensional effort t = (t₁, t₂, …, t_n) across several distinct activities or dimensions of performance, Holmström and Milgrom showed that incentives do not merely induce total effort and allocate risk—they also direct the allocation of attention across competing tasks. When tasks are technological substitutes in the agent's cost function (C₁₂ > 0) and differ in the precision with which they can be measured, providing high-powered piece-rate incentives on an easily measured task (such as output volume or standardized test scores) induces the agent to reallocate effort away from poorly measured tasks (such as product durability, customer goodwill, or conceptual learning). This mechanism yields the Exclusion Principle (explaining why salaried employment with zero piece rates is optimal even when a low-risk-aversion agent performs a measurable activity), provides a unified agency-theoretic boundary between independent contracting and employment (asset ownership), and establishes formal rules for job design by separating easily measured tasks from hard-to-measure tasks across specialized workers.

Classical one-dimensional moral hazard models (Mirrlees, 1975; Holmström, 1979; Grossman and Hart, 1983) evaluate a trade-off between risk sharing and effort incentives for a single scalar effort choice t ∈ ℝ₊. In the single-task linear-normal-CARA benchmark (Holmström and Milgrom, 1987), a risk-neutral principal facing an agent with constant absolute risk aversion r > 0, quadratic effort cost C(t) = (c/2)t², and noisy output x = t + ε with variance σ² sets the optimal commission slope equal to:

Although this single-task formula explains why noisy performance measures (σ² > 0) and worker risk aversion (r > 0) reduce b^* below 1, it fails to account for two empirical regularities in organizations:

1. Pervasiveness of Fixed Salaries (b = 0): Across corporate, bureaucratic, and professional jobs, employees receive fixed time-based salaries or low-powered bonuses (b ≈ 0), even when objective physical counts of output (such as lines of code written, patients seen per hour, or legal motions filed) are nearly costless to record (σ₁² ≈ 0). Under the single-task formula, σ₁² → 0 implies b^* → 1 regardless of risk aversion r.
2. Co-Determination of Compensation, Ownership, and Authority: Moving an activity inside the firm (transforming an independent contractor who owns the productive asset into an employee who uses firm-owned equipment) is systematically accompanied by a shift from high-powered market piece rates to low-powered fixed wages and bureaucratic work rules.

Holmström and Milgrom (1991) resolved both anomalies by building on their 1987 continuous-time Brownian motion aggregation theorem, which proved that an agent with constant absolute risk aversion (CARA) who controls the drift of a multidimensional Brownian motion over time responds to any end-of-period reward scheme as if facing a linear compensation contract.

Technology, Signals, and Preferences

Consider a risk-neutral principal who contracts with a risk-averse agent over a vector of n non-negative task efforts t = (t₁, t₂, …, t_n)′ ∈ ℝ₊ⁿ:

• Principal's Gross Benefit: Agent efforts generate expected gross value B(t) for the principal, where B: ℝ₊ⁿ → ℝ is concave and differentiable with gradient B′(t) = (∂B/∂t₁, …, ∂B/∂t_n)′ (in the linear-benefit case, B(t) = B′t).
• Agent's Multidimensional Cost of Effort: Exerting t imposes monetary-equivalent disutility C(t) on the agent, where C: ℝ₊ⁿ → ℝ is twice continuously differentiable and convex, with gradient C′(t) and positive-definite n × n Hessian matrix C″(t) = [∂²C / ∂t_i∂t_j].
• Observable Performance Signals and CARA Utility: The principal observes n verifiable signals x = μ(t) + ε = t + ε, where ε ~ N(0, Σ) has covariance matrix Σ with diagonal error variances σ_i² = Var(ε_i) and covariances σ_ij = Cov(ε_i, ε_j). If a task k is unmeasurable, σ_k² = ∞. Under the linear wage schedule w(x) = α + b′x = α + ∑_i=1ⁿ b_ix_i, the agent has CARA utility U(w, t) = −e^{−r [w(x) − C(t)]} with absolute risk aversion r > 0.

Certainty Equivalent and Derivation of the Optimal Incentive Vector

Because ε ~ N(0, Σ), the agent's certainty equivalent (CE) wealth is CE_A(α, b, t) = α + b′t − C(t) − (1/2)*r*b′Σb, and the risk-neutral principal's expected payoff is CE_P(α, b, t) = B(t) − α − b′t. Any Pareto-efficient contract maximizes the total joint certainty equivalent surplus S(t, b) = CE_A + CE_P:

For an interior solution t ≫ 0, the agent's first-order condition equates the piece rate on each task i to its marginal effort cost: b = C′(t), with Jacobian ∂t / ∂b = [C″(t)]⁻¹. Differentiating S(t(b), b) with respect to b yields [∂t/∂b][B′(t) − C′(t)] − *r*Σb = [C″]⁻¹[B′ − b] − *r*Σb = 0, which rearranges into the Holmström–Milgrom Multitask Optimal Incentive Formula:

Consider the two-task benchmark (n = 2) with linear benefits B(t₁, t₂) = B₁t₁ + B₂t₂ and independent measurement errors (σ₁₂ = 0, so Σ = diag(σ₁², σ₂²)). Let C_ij ≡ ∂²C / ∂t_i∂t_j denote the entries of the cost Hessian C″ (with C₁₁ > 0, C₂₂ > 0, and Δ_C = C₁₁C₂₂ − C₁₂² > 0).

Effort Substitution vs. Effort Complementarity

From ∂t / ∂b = [C″]⁻¹, the responsiveness of efforts (t₁, t₂) to piece rates (b₁, b₂) is:

• Substitutes in Effort Cost (C₁₂ > 0): Raising effort on task 1 increases the marginal disutility of effort on task 2 (for instance, when both tasks compete for the agent's finite working hours or attention, as in C(t₁ + t₂) where C₁₁ = C₂₂ = C₁₂ = c″ > 0). Here ∂t₂ / ∂b₁ = −C₁₂ / Δ_C < 0: raising the piece rate b₁ on task 1 crowds out effort t₂ on task 2.
• Complements in Effort Cost (C₁₂ < 0): Raising effort on task 1 lowers the marginal cost of task 2, so ∂t₂ / ∂b₁ > 0.

Solving (I + r C″ Σ) b^* = B′ explicitly for b₁^* when σ₁₂ = 0 yields:

Asymmetric Measurability and "Teaching to the Test"

Suppose task 1 (quantity, speed, or standardized drill scores) is measured with finite error variance σ₁² < ∞, whereas task 2 (product quality, safety, or creative critical-thinking instruction) is unmeasurable (σ₂² → ∞, which enforces b₂^* = 0). Taking the limit as σ₂² → ∞ yields the constrained optimal piece rate on the measurable task 1:

This equation delivers three core insights:

1. Even with Noiseless Measurement (σ₁² = 0) or Risk Neutrality (r = 0), Piece Rates Are Muted: When σ₁² = 0 or r = 0, the optimal piece rate on task 1 is b₁^* = B₁ − B₂(C₁₂ / C₂₂) < B₁ whenever C₁₂ > 0, reducing the opportunity cost of devoting effort to the unmeasured task t₂.
2. Low-Powered Incentives as Distortion Control: In public education, if teachers allocate time between drill-based test preparation (t₁, measured by standardized exams) and fostering analytical reasoning (t₂, poorly measured in the short run), high-powered merit pay tied to test scores (b₁ ≫ 0) causes teachers to "teach to the test"—reallocating classroom hours toward t₁ at the expense of t₂.
3. The Attention-Budget Exclusion Theorem: Suppose the agent's cost function depends only on total working time T = t₁ + t₂, so C(t₁, t₂) = c(t₁ + t₂) with C₁₁ = C₁₂ = C₂₂ = c″ > 0. If task 2 is unmeasurable (b₂ = 0), any positive piece rate b₁ > 0 on task 1 induces a corner solution where unmeasured effort t₂ = 0. To induce positive effort on both tasks (t₁ > 0 and t₂ > 0), the principal must set b₁^* = b₂^* = 0 (a fixed salary).

Holmström and Milgrom (1991) used this multitask apparatus to explain firm boundaries and asset ownership, complementing both transaction cost economics and the Grossman–Hart–Moore property-rights theory. Suppose an agent performs two activities using a productive capital asset (such as a delivery truck or manufacturing machine):

1. Direct Production (t₁): Produces verifiable current output x₁ = t₁ + ε₁ (noise variance σ₁²) that accrues to the principal.
2. Asset Maintenance and Care (t₂): Preserves the terminal value of the capital asset, v = t₂ + ε₂ (high noise variance σ₂²). Asset value v is non-contractible via explicit performance pay; whoever legally owns the asset receives its terminal residual value v (and bears its shock ε₂).

This institutional structure creates two coupled governance regimes:

• Independent Contracting (Agent Owns the Asset): The contractor receives 100% of the asset's terminal value v = t₂ + ε₂, fixing the incentive weight on asset care at b₂^IC = 1 (and bearing asset risk (1/2)*rσ*₂²). Because t₁ and t₂ are substitutes (C₁₂ > 0) and b₂^IC = 1 is high, the principal offers the independent contractor a high-powered output price b₁^IC = (B₁ + rC₁₂σ₁²) / (1 + rC₁₁σ₁²) on current production x₁ so the contractor does not neglect deliveries t₁.
• Inside Employment (Principal / Firm Owns the Asset): The firm absorbs the asset-value variance σ₂², insuring the risk-averse worker, and the worker's direct return to asset care is b₂^Emp = 0. Because the employee has zero residual claim on the asset's resale value (b₂^Emp = 0), giving the employee a high-powered piece rate b₁ on output speed would induce the employee to overdrive and abuse the firm's equipment (t₂ → 0). Consequently, when the firm retains ownership of the asset, the firm simultaneously lowers the worker's output piece rate to b₁^Emp < b₁^IC (paying a muted bonus or fixed salary).

Thus, high-powered output incentives (b₁ high), agent asset ownership (b₂ = 1), and contractor autonomy form one coherent cluster (market contracting), whereas low-powered wage incentives (b₁ ≈ 0), firm asset ownership (b₂ = 0), and managerial work rules form an opposing coherent cluster (inside employment).

Optimal Task Bundling Across Multiple Agents

When dividing tasks k ∈ {1, …, N} with measurement error variances σ_k² between two specialized workers, Agent A and Agent B, optimal job design groups tasks according to the homogeneity of their measurement precision. If a single employee is assigned both a cleanly measurable routine task (σ₁² small) and a noisy creative or quality-assurance task (σ₂² large), the Exclusion Principle forces the firm to mute incentives on both tasks. By splitting the tasks into two separate specialized jobs—assigning the easily measured task to Agent A under high-powered commission pay (b_A^* high) and the hard-to-measure task to Agent B under a fixed salary (b_B^* ≈ 0)—the firm achieves high effort on the measurable task without distorting attention away from the unmeasured task.

Bureaucratic Rules and Congruence Extensions

Let task t_out represent an outside personal business activity ("moonlighting") that the agent can pursue during the workday (b_out = 1 if permitted, B_out = 0). If the employer institutionally prohibits outside activities during working hours (setting t_out ≡ 0 via attendance monitoring and exclusive employment covenants), the opportunity cost of the employee's time on the job falls, allowing the employer to motivate internal tasks (t₁, t₂) using low-powered salaries (b₁, b₂ small).

Baker–Gibbons–Murphy Performance Alignment

George Baker (1992) and Baker, Gibbons, and Murphy (1994) extended the multitask framework to misaligned performance measures: when the principal's true objective is y = f′t + ε_y (non-contractible) while the verifiable metric is p = g′t + ε_p, the optimal bonus slope on p depends on the geometric cosine congruence between the vectors f and g in the metric of [C″]⁻¹, often requiring subjective bonus assessments sustained via repeated-game relational contracts.

How does the Holmström–Milgrom (1991) Multitask Principle modify Holmström's (1979) Informativeness Principle?

In Holmström's 1979 Informativeness Principle (derived for a single scalar effort t), adding any costless verifiable signal x₁ that contains information about t never harms the contract. In the multitask environment (1991), conditioning pay on an accurate signal x₁ of task t₁ when a competing substitute task t₂ is unmeasurable (σ₂² = ∞) can reduce total welfare relative to ignoring x₁ (b₁ = 0), because crowding out t₂ outweighs the benefit of motivating t₁.

Why does the multitask model predict that piece rates can be zero even when output is measured without noise (σ₁² = 0) and the agent is risk-neutral (r = 0)?

In the multitask model with substitute tasks (C₁₂ > 0) and an unmeasurable second task (b₂ = 0), setting b₁ > 0 distorts the relative marginal reward between task 1 and task 2 (b₁ vs. 0). Even if r = 0 and σ₁² = 0, the formula b₁^* = B₁ − B₂(C₁₂ / C₂₂) depends on the substitution ratio C₁₂ / C₂₂; when C₁₂ = C₂₂ and B₂ ≥ B₁, the optimal piece rate is b₁^* = 0.

How does the Holmström–Milgrom explanation of asset ownership differ from the Grossman–Hart–Moore property-rights model?

In the Grossman–Hart–Moore property-rights model, all parties are risk-neutral, no explicit performance contracts can be written ex ante, and asset ownership shifts ex-post bargaining threat points. In Holmström and Milgrom (1991), explicit linear performance contracts (b₁) are available, the agent is risk-averse, and asset ownership bundles the residual return and risk of non-contractible asset maintenance (b₂ = 1 vs. b₂ = 0) with explicit wage incentives (b₁) and workplace authority rules.

1. Holmström, B., & Milgrom, P. (1991). "Multitask Principal–Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design." Journal of Law, Economics, & Organization, 7(Special Issue), 24–52. https://doi.org/10.1093/jleo/7.special_issue.24
2. Holmström, B., & Milgrom, P. (1987). "Aggregation and Linearity in the Provision of Intertemporal Incentives." Econometrica, 55(2), 303–328. https://doi.org/10.2307/1913238
3. Holmström, B., & Milgrom, P. (1994). "The Firm as an Incentive System." American Economic Review, 84(4), 972–991. https://doi.org/10.1257/aer.84.4.972
4. Baker, G. P. (1992). "Incentive Contracts and Performance Measurement." Journal of Political Economy, 100(3), 598–614. https://doi.org/10.1086/261831
5. Baker, G., Gibbons, R., & Murphy, K. J. (1994). "Subjective Performance Measures in Optimal Incentive Contracts." Quarterly Journal of Economics, 109(4), 1125–1156. https://doi.org/10.2307/2118358