You are on page 1of 27

Incentives to Help in Multi-Agent Situations Author(s): Hideshi Itoh Source: Econometrica, Vol. 59, No. 3 (May, 1991), pp.

611-636 Published by: The Econometric Society Stable URL: http://www.jstor.org/stable/2938221 . Accessed: 11/09/2013 08:41
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.

http://www.jstor.org

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

Econometrica, Vol. 59, No. 3 (May, 1991), 611-636

INCENTIVES TO HELP IN MULTI-AGENT SITUATIONS BY HIDESHI


ITOH1

This paper concerns moral hazard problems in multi-agent situations where cooperation is an issue. Each agent chooses his own effort, which improves stochastically the outcome of his own task. He also chooses the amount of "help" to extend to other agents, which improves their performance. By selecting appropriate compensation schemes, the principal can design a task structure: the principal may prefer an unambiguous division of labor, where each agent specializes in his own task; or the principal may prefer teamwork where each agent is motivated to help other agents. We provide a sufficient condition for teamwork to be optimal, based on its incentive effects. We also show a nonconvexity of the optimal task structure: The principal wants either an unambiguous division of labor or a substantial teamwork.
KEYWORDS: Principal-agent relationships, moral hazard, multiple tasks, team production, incentives to help.

1. INTRODUCTION

THIS

PAPER CONCERNS moral hazard problems in multi-agent situations where cooperation is an issue. We consider a situation where each agent can allocate his effort to various production activities called tasks. Tasks are assumed to be "independent" of each other: the outcome of each task depends on an exogenous random variable which is stochastically independent of the random variables affecting the other tasks; and revenues from each task only depend on the outcome of that task. Relative performance evaluation therefore does not give a reason for the wage schedule to an agent to depend on the outcome of the tasks assigned to the other agents. (See Baiman and Demski (1980), Green and Stokey (1983), Holmstrom (1982), Lazear and Rosen (1981), Mookherjee (1984), and Nalebuff and Stiglitz (1983).) We focus on incentives of agents to "help" each other. Each agent chooses his own effort level which improves stochastically the outcome of the task for which he is mainly responsible. Agents also choose the amount of "help" to extend to other agents which improves the outcomes of their tasks. The principal, who cannot observe the effort chosen by each agent, designs wage schedules contingent on outcomes. By selecting appropriate wage schedules, the principle can design a task structure: The principal may prefer a specialized task structure, where each agent is inclined not to help other agents and specializes in his own task. In this case, by the assumption of independent tasks, each agent can be treated completely separately. The principal however may choose a nonspecialized task structure, called teamwork, in which agents are motivated
1

This paper is based on a chapter in my doctoral dissertation submitted to Stanford University, March, 1988. I am grateful to Masahiko Aoki for inspiring me to this research, to David Baron, George Mailath, John McMillan, Dilip Mookherjee, Mark Wolfson for helpful comments, and especially to a co-editor and two anonymous referees for their very careful comments as well as editorial assistance, and to David Kreps for his constructive suggestions and constant encouragement. 611

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

612

HIDESHI ITOH

to help each other. In this second case, the principal has to make the wage contract to an agent contingent on the outcomes of the other agents' tasks. Since the outcome of each task is affected by both the own effort of the agent assigned to that task and the helping effort by the other agents, this nonspecialized task structure leads to team production.2 In our model, team production endogenously appears as a result of the wage schemes chosen by principal. This research should be distinguished from the previous literature that analyzes how disadvantages of team production, namely the unobservability of the marginal contribution of each agent and the resulting free-rider problem, are resolved in situations where team production exogenously exists because of the nature of tasks or production processes (Alchian and Demsetz (1972), Holmstrom (1982), McAfee and McMillan (1986)). Situations where each agent allocates effort to multiple tasks have been analyzed by Drago and Turnbull (1987, 1988) and Lazear (1989). In these papers particular forms of compensation schemes are exogenously imposed. Drago and Turnbull (1987) and Lazear (1989) consider tournaments and show that when agents play effort choice subgames noncooperatively, positive helping effort does not occur in equilibrium. In addition, Drago and Turnbull (1987) show that this is also true under noncompetitive promotion schemes such as a quota scheme, and then they consider cases where agents can reciprocate via binding agreements. Lazear focuses on "negative helping effort," called sabotage, and shows that pay compression between the winner and the loser may be preferable in order to reduce sabotage by agents. Drago and Turnbull (1988) examine whether agents efficiently allocate given total effort between own effort and helping effort under two particular linear compensation schemes; an individual piece rate scheme and a group piece rate scheme. They find that depending on whether the agents behave noncooperatively or cooperatively, individual piece rate schemes may lead to "inefficient under-cooperation," while group piece rate schemes may cause "inefficient over-cooperation." Rather than specifying some particular form of wage schedules, I follow the literature in the "standard" agency models with hidden action (but without hidden knowledge), in particular, Grossman and Hart (1983), Mookherjee (1984), and Rogerson (1985), and examine the general incentive problem in the situation explained above.3 Although this approach usually does not yield many sharp predictions, the logic behind the problem will be clarified. We analyze a hidden action model of the relationship between a principal and two risk averse agents who select efforts noncooperatively.4 Because of the independence across tasks, one would only find in the model that the wage schedule to an agent is contingent on the outcomes of the two tasks in order to give him an
2According to Alchian and Demsetz (1972), team production exists when several inputs to a production process are utilized by more than one individual and the product is different from the sum of the outputs from inputs separately used by each individual. 3Recently, Holmstrom and Milgrom (1989) have started to study related issues of multitask agents in the linear model developed by Holmstrom and Milgrom (1987). 4 The case where the agents can collude via side contracts is analyzed in Itoh (1990).

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

613

incentive to help the other agent: When the principal wants neither agent to choose positive helping effort, it is strictly better to treat them completely separately. The main purpose of this paper is to examine what factors affect the principal's decision to induce team production through interdependent incentive schemes rather than to treat the agents separately through individual-based schemes. To this end, we utilize the first-order approach5 and examine how a marginal change of the optimal independent contract in the direction of teamwork, by making the payments to an agent contingent on the outcome of the other agent's task as well, affects the principal's welfare. There are two questions associated with this. First, is such a marginal change in fact able to induce an agent to help the other agent? Second, if so, does this actually work to the principal's advantage? Or could it be that help has undesirable incentive effects on the other agent that outweigh whatever direct advantages it might have? The answer to the first question is yes when the agent's marginal disutility with regard to helping effort is zero at zero help. This case is likely to apply if agents have task specific preference. Teamwork is then always optimal when the problem is risk sharing only. However, the existence of the moral hazard problem and the resulting strategic interaction among agents may lead the principal to adopt individual-based schemes because teamwork actually works to his disadvantage. We provide, as an answer to the second question posed above, a sufficient condition under which this is not the case and teamwork actually is optimal. The condition represents the effects of introducing "a small amount of teamwork" on the principal's welfare through the incentive compatibility constraints. In particular, teamwork is optimal if own effort and helping effort are complementary so that an agent responds to an increase in help from the other agent by increasing his own effort. Even in cases where an agent's optimal response to an increase in help is to reduce his own effort (the case of free-riding), teamwork is optimal if the resulting decrease in own effort reduces the costs of inducing him to work appropriately on his own task sufficiently. Such a case holds when an agent's task is so monotonous that the marginal productivity of own effort decreases or the marginal disutility of own effort increases drastically as he raises his own effort level. If tasks are similar and agents only care about the total amount of effort, they are reluctant to provide even a small amount of help because the marginal disutility of helping effort at zero help is positive. Then the answer to the first question turns out to be no. That is, perturbing the optimal independent wage schedule in the direction of teamwork by a small amount cannot elicit positive helping effort. A large perturbation is required to induce any help, and with risk averse agents this has a substantial cost in terms of inefficient risk sharing. Put
5Unlike in the one-agent model, the monotone likelihood ratio property (MLRP) and the convexity of the distribution function condition (CDFC) are not sufficient for the first-order approach to be valid in our model. We will present a sufficient condition which consists of MLRP, a generalization of CDFC, and a condition on the agents' utility functions for income.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

614

HIDESHI ITOH

differently, if we decompose the problem as is done in Grossman and Hart (1983), the cost function for implementing various levels of effort increases discontinuously as we move from zero to positive help. Since the expected revenue function is continuous, a little bit of help will always be suboptimal. Thus, we fail to obtain sufficient conditions for teamwork to be optimal by examining the local necessary conditions which must be satisfied by the optimal independent contract. However, this may partially explain why the personnel management of the firm is extreme in the sense that we observe either a strict division of labor through individual wage schemes or substantial teamwork through team-based wage schemes. Note that the existence of multiple agents is not at all necessary for this nonconvexity result. The result applies whenever a single agent has options of allocating his effort to multiple tasks. While the paper focuses on theoretical issues of incentive problems due to hidden action of the agents working on multiple tasks, the central question asked in the paper is discussed in various recent literature in economics and management: U.S.-Japan comparisons in organization structures and personnel policies (Aoki (1988), Kagono et al. (1985), Lincoln and McBride (1987)); management of American high performance companies (Waterman (1987)); work design for individuals and for groups (Hackman and Oldham (1980)); a recent research on productivity performance in American manufacturing industries (Dertouzos et al. (1989)). The current paper is intended to be a step toward the economic analysis of teamwork and optimal task design. The rest of the paper is organized as follows. In Section 2, the model is presented. Some preliminary results are obtained in Section 3. Sections 4 and 5 are the main part of the paper. In Section 4, a sufficient condition for teamwork to be optimal is given. The nonconvexity result is presented in Section 5. Section 6 is concluding remarks.
2. THE MODEL

We consider the relationship between one principal and two agents n = 1, 2. Extension to the general case of N agents is straightforward, and will be briefly discussed in the final section. Since task assignment is not our concern, we assume that agent n is exogenously assigned to a technologically well-defined, "independent" task n because, for example, only he has some necessary expertise for that task or both agents are equally capable. Task n is independent of task k (k 0 n) in the sense that the outcome of task n, which is a random variable, is stochastically independent of the outcome of task k.6 For simplicity, we assume that the outcome of each task is either success (S) or failure (F). Let pf be the probability of success in task n, and Pi' be the joint probability that the outcomes of task n and task k are i and j, respectively (i, j E {S, F}). For example, PsF =p'(1 _pk) Each agent simultaneously chooses a level of effort, which jointly determines the probability distribution of the outcomes of the tasks. The effort level chosen
6

Unless otherwise noted, whenever n and k appear, it is assumed that k 0 n holds.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

615

by agent n is a two-dimensional variable en = (an, bn)E &x _ [0, A] x [0, B], where an is the effort expended on his own task, called own effort, and bn is for the task occupied by agent k, called agent n's helping effort for agent k.7Agent n's own effort and agent k's helping effort jointly affect the probability distribution of the outcome of task n: the success probability of task n is a function of the input (an, bk), and is written as pn(an, bk). The joint probability Pij is hence the function of (en, ek) and is written as Pi (en, ek). We assume that for each n, pn(.) is twice continuously differentiable with regard to its arguments, and denote the partial derivatives by pn, Pnb' and so on. We adopt the following assumptions on pn(.).
ASSUMPTION1: For each n, (i) pn(a, b)> 0 for all (a, b); (ii) pn(a, b)> 0 for all (a, b), with strict inequality if a > 0; (iii) pn(O, 0) > pk(0, 0); and (iv) pn(a, b) E (0, 1) for all (a, b).

Part (ii) gives the reason b is called helping effort. It is plausible to assume that agent k, who is not assigned to task n, can increase the productivity of task n through various kinds of help. Agent k's doing some routine work for agent n may help the latter improve his productivity by concentrating on more sophisticated work in task n. Or agent k may possess skills or knowledge which, while not essential to task n, are still valuable at that task. Note that (i) and (ii) are equivalent to the strict monotone likelihood ratio property (strict MLRP) in our two-outcome model. Thus by Milgrom (1981), observation of success in task n is "more favorable than" observation of failure: a helping effort by agent k (or an own effort by agent n) fixed, the posterior distribution of the own effort of agent n (the helping effort of agent k, respectively) given observation of success in task n dominates the posterior probability distribution given observation of failure in the sense of the first-order stochastic dominance. Part (iii), with an assumption on agents' utility functions stated shortly, excludes the case where the input into a task is only the helping effort from the agent assigned to the other task. One extreme, but possible case in which this assumption holds is that agent n's own effort is required in order for the other agent's help to improve the productivity of task n, that is, pg(0, b) = 0 for all b. The final part (iv) states that there is no moving support. Let Tij be the total revenue earned by the principal when the outcomes of task 1 and task 2 are i and j, respectively. The expected total revenue, denoted by R(e1, e2), is then given as R(el, e2) = FEijPi1(e1, e2)wr1.To focus on endogenous formation of team production, we adopt an inessential assumption that the principal'srevenue is additivelyseparablein tasks: Tij = ,T + 7j where Tin is the revenue from task n when its outcome is i = S, F with wTr'> w'. The expected total revenue of the principal is then written as R(e1, e2) = Hl(a1, b2) + H2(a2, bl) where H' is the expected revenue from task n defined by
H'(an,a bk) = wF + p (an,a bk )( 7n)

7 For simplicity, we take A > 0 and B > 0 sufficiently large so that we ignore the case where effort levels reach these upper bounds.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

616

HIDESHI ITOH

The principal selects a wage schedule for each agent. She (the principal) has the ability to commit herself to that schedule. The outcome of each task is assumed to be publicly observable, so that the wage schedule can be contingent on the outcomes of the two tasks. Denote by wj the wage paid to agent n when the outcomes of task n and task k are i and j, respectively. The wage schedule for agent n is a four-dimensional vector w' = (w's, W'F, wF, wFF). The (von Neumann-Morgenstern) utility function of agent n is defined on V"x Mx 97 where V" is an interval (w,oo). We further assume that the utility function is twice continuously differentiable and has the additive form Vn(w)- Gn(en) where Vn is strictly increasing and (weakly) concave. That is, the agents are assumed to be risk neutral or risk averse. Concerning the disutility of effort, we and so on. The assumptions on these denote the partial derivatives by Gan, Ganb, derivatives are as follows:
ASSUMPTION 2: For each n, (i) GQn(a,b) > 0 for all (a, b), with strict inequality if a > 0; (ii) Ggn(a, b)> 0 for all (a, b), with strict inequality if b > 0; (iii)

Gan(O,0) < Gn(0, 0); and (iv) Gn(v)

is strictly convex.

Assumption 2 states that the agents are generally effort averse and that their disutility rises at an increasing rate as they work harder. And in order to induce agents to exert positive own effort levels, it is assumed that when neither own effort nor helping effort is positive, the marginal disutility of helping effort is at least as large as that of own effort. The most important part of Assumption 2 is that given a positive own effort, we allow a case where a marginal disutility of helping effort at zero help is strictly positive as well as a case where it is zero. The former case occurs, for example, when an and bn represent the time agent n allocates to his own task and the other task and the disutility of work depends only on the total amount of time he works; Gn(a, b) = Gn(a + b). Then given a positive own effort, a small amount of help increases the agent's working time and increases his disutility of effort. The latter case may occur when the agents have task specific disutility for each task because, for example, tasks are of very different types. An example is Gn(a, b) = Kn(a) + L (b) with K,I(O) = LI (O) = 0. For the wage schedule wn and the effort combination (en, ek), let Un(Wn, en, ek) be the expected utility of agent n, defined by Un(wn, en, ek) = - Gn(en). The principal is assumed to be risk neutral and EE1Pijn(en, ek)Vnf(w;j) to choose wage schedules (w', w2) and effort levels (el, e2) to maximize her expected residual profits. The principal's Original Problem, denoted by (OP), is given as follows: max
w ,w ,ei,e2

R(e1,e2)

EPi(el,e2)(w11j+w
j

subject to

(OP)

(NIC) Un(Wn, en, ek) > Un(Wn, el, ek)


(PC) U (w , en, ek) > Uj for n = 1,2,

for all el and n = 1,2,

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

617

where U' is the reservationutility level of agent n. The constraints(NIC), constraints,state that, givenwage schedules, called Nash incentivecompatibility the effort choice of the agents forms a Nash equilibrium.The last constraints constraints. (PC) are the participation FollowingGrossmanand Hart (1983), we can decomposeproblem(OP) into Problem which solves, given effort levels two parts: (i) the Implementation (e1,e2), the wage schedules that minimizethe expected wage paymentsof the principal to the agents subject to (NIC) and (PC); (ii) the Effort Selection Problemwhich finds the effort levels that maximizethe principal'sexpected net profits. Formally, these problems are described as follows. Let v1'= Vn(wj) and
=n(J'().

Instead of wn, we can regard

each n as the choice variableof the principal.We sometimescall Vn(as well as wn) a wage schedule for agent n. Let hn be the inverse function of the utility function on income Vn.Then if the principalwants to implementeffort levels (e1, e2), she has to solve the followingproblem(IP): (IP) min
v,V i

SS=(v,

VSF, VF,S VFF)neY

for

EPi(e1,
J

e2)(h1(v)

+ h2(vj))

subjectto (NIC) and (PC).

Note that all the constraintsare linear in v/. Thus, if agents are risk averse so that hn is strictlyconvex for all n, problem (IP) for each effort combination problemwith an infinite numberof (e1, e2) is a standardconvex programming linear constraints.If the feasible set in (IP) is nonempty,the solution(v1, v2) to (IP) exists. This can be proved by following Grossmanand Hart (1983) and Mookherjee(1984).We call the solutionthe optimalwage schedulesfor (e1, e2).
Let C(el,
e2)

be the minimum cost to implement (e1, e2), which is the optimal

value of problem (IP) for that effort pair. Then the Effort Selection Problem (EP) solves (EP)
maxR(e1, e2) - C(e1, e2).
el,e2

The existence of the solution to (EP), called the second-best efforts, can be shown by using argumentsanalogous to those in Grossmanand Hart (1983). Throughoutthe paper, it is assumed that the second-best efforts are not the least costlyones (en = (0, 0) for all n). In addition,we supposethat the principal wants each agent to work at least on his own task.
ASSUMPTION

3: The solution to (EP) satisfies an

>

O for n = 1,2.

We make this assumptionin order to simplifythe exposition in the paper, in Section 4, as well as to focus on incentives of mutual help: it particularly excludesthe case in which the principalwants both agents to workonly on one task (a2 = b1 = 0 or a1 = b2 = 0). However, the results in this paper hold without are requiredas we will commentlater only minormodifications this assumption: in Section4. One sufficientconditionfor Assumption3 to be true is G n(0, b) = 0

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

618

HIDESHI ITOH

for all b and n, because then the marginalbenefit of increasingown effortfrom zero outweighsthe marginalcost, regardlessof help, and hence the principal wants each agent to work at least a little bit on his own task.
3. PRELIMINARY RESULTS

As a benchmark for the analysis,we define the first-bestsolutionwhich is the effort combinationthe principalwants to implementwhen she can observe the effort choice of each agent. To ensure that the principalcan implement any effort level by some wage schedule that guarantees the agent exactly his reservationutilitylevel, we assumethe following:8
ASSUMPTION

4: For each n, (i) limw w Vn(w)=

oo;and (ii) Un +

Gn(en)E

for all en.

Define the first-bestcost function CnFB(*) for agent n by CFnB(en) = h Gn(en)).The first-bestsolution solves the followingproblem(FB): (FB)
maxR(el,
el,e2

(Un +

e2) -CF1B(elj) -CFB(e2).

be the first-besteffortof agent n. A few characteristics Let en = (an*bn*) of the first-bestsolutionwill be found in the next section. In particular, we will obtain a necessaryand sufficientconditionfor the first-besthelping effort for an agent to be positive,by examiningfirst-order conditions. In the second-bestsituationwhere the effortof each agent is noncontractible, the principalutilizes the observation of outcomesin order to provideeach agent with incentivesto choose desirableeffort levels. To induce each agent to work on his own task, his wage schedulemustbe contingenton its outcome.Similarly, the wage to agent n must depend on the outcome of task k if the principal wants agent n to choose a positive helping effort for agent k. Generally,this incentiveeffect distortsthe optimalrisksharingbetween the principaland each agent, and thereby the principal incurs a loss due to the unobservability of effort. As in the standardone-agentmodel with moralhazard,however,when all the agents are risk neutral,there is no such loss and the principalcan achieve the first-bestsolution.This is obviouswhen the first-bestsolution has zero help for both agents. The principal then has an independent relationshipwith each agent, and hence she can induce the first-bestown effort by payingeach agent the whole marginalrevenuefromhis task. Note that the principalcan attainthis without bearing any risk by subtractingthe expected net profit from task n, Hn(a,*, O)- CnB(a*, 0), from the payment to agent n. The argumentextendsto the case in which the first-bestsolution has positive > 0. The principalis now able to induce the help for some agent. Suppose bn*
8Assumption

4 (ii) holds automatically if the agents' utility-of-income functions Vn are un-

bounded above.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT SITUATIONS

619

first-besteffort by payingthe marginaltotal revenuesfrom both tasks to agent


n. And she subtracts the fixed amount R(e , e2 ) - CnB(e*) from the payment.9

Under this scheme, however, the principalpays the whole marginalrevenue from task k to both agents,and therebyshe prefersthe outcome of task k to be failureratherthan success.The principalhence has to share some riskwith risk neutral agents.10
PROPOSITION 1: Suppose that all the agents are risk neutral. Then the principal can achieve the first-best solution by an interdependentlinear scheme which pays each agent the whole marginal increases of the revenuesfrom all the tasks.

Now suppose that at least one of the agents, say, agent n, is risk averse. If the

principalwants him to select positive help, his wage schedule must depend on the outcome of task k. The converseis also true:If the helpingeffortof agent n is zero, then the optimalwage schedule should be independentof the outcome of task k. This is an applicationof the so-called"sufficient statisticstheorem"in multi-agentsituationsby Holmstrom(1982) and Mookherjee(1984). This result providesthe sufficientand (partial)necessaryconditionfor optimalpaymentsof an agent to be independent of the other agent's performance.The sufficient
condition for agent n's contract is that given ek' for all en, een,iJ(en edl e^k) is independent of j (i, j = S, F). That is, the outcome of task n is a P1nj(e,,

sufficientstatisticfor the effort choice by agent n. In our model, this ratio is of the form
Pin(an, bk)Pjk ak S bn)

(1)

Pin(an, bk) j(ak, bn)


pf(.)

where Pn( )

1 pfn( ). When our attention is confined to the subset of feasible efforts in which help is equal to zero (bn = bn= 0), (1) is wage scheme to agent n is better independentof j. Thus, the individual-based from the risk sharingpoint of view. and Pp(.)
=

PROPOSITION 2: Suppose that agent n is risk averse. Then the optimal wage schedule of agent n is contingent on the outcome of task k if and only if bn > 0
9 By following McAfee and McMillan (1986), we will be able to generalize the optimality of the similar linear wage schedule to the case where hidden knowledge as well as hidden action exists. When hidden knowledge exists, because of the information rents due to private information held by the agents, the first-best solution is not achievable. And the principal does not give the agents all of the marginal increases in the total profits in order to extract some of their information rents. 10One might ask whether there is any other wage scheme implementing the first-best solution with no risk imposed on the principal. The answer is no, at least in our model of two possible outcomes. The proof is presented in earlier versions of the paper. The intuition goes as follows. In order to induce the agents to choose the first-best inputs to task k, the principal has to pay the whole marginal expected revenue from that task to each agent. However, when the principal gives agent n the whole marginal expected revenue from task k and she does not incur any risk, the marginal expected revenue from task k paid to the other agent must be zero. Then the principal cannot provide agent k with an incentive to choose a positive effort for his own task. She therefore must bear some risk when she wants to induce help even though the agents are risk neutral. This contrasts with the case of a relationship with one agent.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

620

HIDESHI ITOH

Proposition2 showsthat the only reason in our model why the wage schedule of a risk averseagent depends on the outcome of the task assignedto the other agent is to give the formeragent an incentiveto help the latter. The reason for the principalto select interdependentwage schemes in our model is therefore differentfrom that in the literatureon relativeperformanceevaluationlisted in Section 1. This literature considers the situation in which the production function is separable in effort and the outcomes of tasks are correlated.The relativeperformanceevaluationis then valuablesince the principalcan use the performanceof an agent as a signal of the effort choice by the other agent. There it is typical to find that the paymentto one agent is decreasingin the other's performance. There are howevercases where the reversewould be true. For example, consider two salesmen attemptingto sell the same goods in a commonterritory.Suppose that there are not manypotential customersin that area so that the outcomesof the salesmenare highlynegativelycorrelatedgiven their effort choice. Then salesman n will be rewardedmore the better is the performanceof salesman k, because a better outcome of salesman k conveys the informationthat salesman n worked harder. On the other hand, in our setting where the outcomes are independent given effort choice, a better outcome of salesman k does not provide any informationconcerningthe own effort of salesman n, but does convey some informationabout salesman n's
helping effort.
4. INDUCED TEAMWORK AS AN OPTIMAL TASK STRUCTURE

The main objective of the paper is to obtain conditions under which the principalmotivatesthe agents to help each other in the equilibrium, that is, the principalendogenouslycreates teamwork.By Proposition1, when both agents are risk neutral,the principalwants them to help one anotherif and only if the first-bestsolution has positive helping effort for both agents. Teamworkin this case therefore does not have any peculiar cost or benefit that does not exist when each agent is paid as a functionof the outcome of his own task only. Thus we assume hereafterthat both agents are risk averse. Suppose that each agent is contracted with independently,with a wage schedule contingentonly on the outcome of his own task. Since there exists no incentive to select positive helping effort under such an individual-based contask structurewhere each tract, we say that the principalchooses a specialized agent specializesin his own task and is not involvedin the other task. Then the question we ask is: under what conditions does a marginal change of the optimalindependentscheme in the directionof teamwork,by makingthe wage schedule of an agent depend on the outcome of the other agent's task as well, welfare? improvethe principal's The analysis uses the first-order approach. Based on Mirrlees (1975), Rogerson (1985) shows that, in the standardone-agent model, the sufficient condition for the first-order approach to be valid is the convexity of the

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

621

distributionfunction condition (CDFC) along with the monotone likelihood ratio property (MLRP). In our model with production externalitiesbetween agents, however,CDFC is not generallysufficient.The problemis that, though we can generalizeCDFC for the joint probabilitydistributionof the outcomes of two tasks,this generalizedcondition,alongwith nondecreasing wage schemes (generally derived from MLRP), does not guarantee the concavity of the expected utility function of the agents. We have to find some sufficientconditions that validatethe first-orderapproachto our model. We define the following Relaxed Problem (RP). Let Un be the partial derivativeof agent n's expected utility Un with regardto its argumentx. Then in (RP), the principalchooses (w1,w2) and (e1, e2) such as a1 > 0 and a2 > 0 to maximizethe objectivefunctionin (OP) subjectto the participation constraints
(PC) and the following local Nash incentive compatibility constraints (LNIC) in

place of (NIC): (LNIC)


Uan(wn, ensek) = 0

and

bnUbn,e(wn,efek)

for n = 1,2.
JF

The next assumption on the joint probability distribution


pk)

=(1-pf)(1-

and the inversefunction hn of Vnturnsout to be sufficientfor validatingthe first-orderapproach.


ASSUMPTION

5: For each n, (i) PFnF(en,ek) is convex in en; and (ii) h ( ) is

convex.

Part (i) is a generalizationof CDFC. From the economic point of view, this conditioncan be interpretedas some kind of stochastically diminishingreturns to scale. Part (ii) can be best understoodby consideringHARA utilityfunctions with - Vn'(w)/Vn"(w)= pw + 0. Then we can show that hn is convex if and only if p < 2. This implies that the coefficient of absolute risk aversion must not decline too quickly.Examplesinclude standardones such as the utilityfunction with a constantabsolute risk aversion(corresponding to p =-I), the logarithmic
utility function (p = 1), and the square root function (p = 2).11
LEMMA 1: For (e1,e2) with an>O, n=1,2, (i) (w',w2) and (e1,e2) solve (RP) if and only if they solve (OP); (ii) the optimal wage schedules are monotone for i =S,F, with the increasing, that is, for each n, w > wFi and w,>wn equality if and only if bn = 0.
1Alternatively we could assume that Psns(en,ek) is concave in en and hl(-) is concave for 1, 2, though utility functions satisfying the latter condition are less standard.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

622

HIDESHI ITOH

The proof is in the Appendix. Lemma 1 makes it possible to analyze the Relaxed Problem instead of the OriginalProblem.'2Because Gn() is strictly convex,Un(Wn *, ek) is also strictlyconcave,given that wn is monotoneincreasing.

situation To make the argumentsimple and clear, we considerthe symmetric in this section:the agents have the same utilityfunctionV(w) - G(a, b) and the same reservationutility U and h = V-1; the success probability of each task is given by the same function p: Q/X 0 -* [0,1]; and the revenue from each task when the outcome of the task is i is wri. For the comparison with the second-bestsolution,we providea necessaryand sufficient condition for positive helping effort to be optimal in the first-best situation.To simplifythe analysisof this benchmarkcase, we assume that the objective function in problem (FB) is strictly concave in (e1, e2). A sufficient condition is that p( ) is concave, which is not implied by Assumption5. This ensures the existence of the symmetricfirst-bestsolution. Let e* = (a*, b*) be the unique first-bestsolutionto problem(FB) and e** = (a**, 0) be the solution to problem(FB) subjectto zero helpingeffort.That is, a** is the first-bestown effort level when the agents are contractedwith independently.
ASSUMPTION

6: The objective function in problem (FB) is strictly concave in

(el, e2).
PROPOSITION 3: Consider the symmetric situation as stated above. Then (i) if a** = 0, the first-best solution e* satisfies a* = b* = 0, that is, e* = e**; (ii) if a** > 0, then a* > 0, and the necessary and sufficient condition for e* =(a*, b*) to satisfy b* > 0 is given by

pb(e

Gb(e**)

Pa( e**)

Ga(e**)

PROOF: The first-orderconditions for a* and b*, which are necessary and sufficient by Assumption 6, are given by PaGiS - ITF) < h'(U + G)Ga and Pb(T7S - WF) ? h'(U + G)Gb, with the equalities if a* > 0 and b* > 0, respectively. Similarly, e** satisfies the first inequality, with the equality if a** > 0. Part (i): If a** = 0, then by Assumptions 1 (iii) and 2 (iii), e** = (0,0) satisfies the first-order condition for b* as well as for a*, and hence e** = e*. Part (ii): Suppose a** > 0. If a* = 0, then b* = 0 must hold. (The argumentis the same

The reason to use the complementarity slackness conditions in (LNIC) is that it allows the solution to (RP) to have zero helping effort: If we used Ub' = 0 instead, the independent wage schedule with zero helping effort could not be a solution in the" relaxed problem. The same problem emerges in the standard one-agent problem: The relaxed problem and the doubly relaxed problem in Rogerson (1985) cannot have the least costly action as an optimal solution, though the implementation of such an action is trivial in his agency model with one-dimensional effort variables. Note also that for the same reason as above, without Assumption 3, we would have to replace the first equation in (LNIC) by the complementarity slackness condition.

12

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

623

as the proof of Part (i).) The least costly effort pair hence becomes the overall first-bestsolution as well as the first-bestsolution subjectto zero help. Contradiction.Thus, a* > 0 must hold. For the necessitypart, suppose Pb/Pa < Gb/IGa at e**. Then by definition,e** satisfiesthe first-order conditionsfor a* and b* given above. This implies that e** is the unique first-besteffort, which contradicts b* > 0. For the sufficiencypart, suppose instead b* = 0. Then by definition, a* =a**, and e** satisfies the first-orderconditions given above. Then, however,Pb/Pa < Gb/Ga at e**, whichis a contradiction. Q.E.D. The intuition is simple. If condition (2) holds, decreasinga small amount of own effort and increasinghelping effort for a task raises the success probability of that task, which effect dominatesthe increasein disutility.Thus, the optimal effort choice has positive helping effort. Conversely,if the condition does not hold, then starting from the first-best independent contract, any marginal increase in helping effort would be less useful than the same increase in own effort,which by the definitionof the first-bestindependentcontractitself is too costly. In particular,if Gb = 0 at zero helping effort,the conditionalwaysholds so that the first-bestsolution is alwaysteamwork.If the disutilityterm depends
only on the total amount of effort a
+

b, (2) becomes

Pb(e**)

> Pa(e**):

the

first-bestsolution is teamworkif and only if helping effort is marginallymore productivethan own effort at the first-besteffort subjectto zero help.
REMARK:

Assumption6 is crucialin proving(i) and the necessitypart in (ii):

without it, one might have a* > 0, b* > 0 while a** = 0; or e** might satisfy the

first-orderconditions while not globally optimal in the unrestrictedfirst-best problem. The sufficiencypart would hold without the assumption,and hence the discussionof local improvements given above would be still correct. Returning to the second-best situation with risk averse agents, denote by (W, wF) the optimal individual-based wage schedule and by eo = (a0, 0) with a0 > 0 the optimal effort levels under that scheme. (The problemis trivial when a0 = 0.) The utility-on-incomelevel correspondingto the wage w? is denoted by vu. While the optimal independentcontract (w0, e0) may not be unique,supposefor simplicity that the principaloffersthe same contractto both agents.The followinglemmaprovidesa necessaryconditionfor offering(wo,eO) to both agents to be the second-bestsolution.The conditionis derivedfrom the Kuhn-Tucker conditionsof problem(RP). For each n, let Anbe the Lagrange multiplier for (PC), ,,n and {n be the multipliersfor the first and second constraintsin (LNIC), respectively.Because of symmetry, these multipliersare
WO=

the same for n = 1, 2, and thereby we denote them by A, ,u, and 6.


LEMMA 2: Consider the symmetric situation.If offering(w0, eo) witha0 > 0 to both agents is the second-best solution, thereexist A > 0, ,u > 0, and 6 > 0 such

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

624
that (3)
A ( pb(eO)
Ap(eO)
-

HIDESHI ITOH

Gb(e )
Ga(e)) Pab(e )

Gb(e0) Ga(e0)
Gab(e?)

Pa(eo)

Ga(e?)

Pb( e?)
Pa( e?)

(Paa(e?)
Pa(eo)

Gaa(eo) Ga(eo)

o0.

The proof is standard. Since (w?,eo) solve (RP), they satisfy the PROOF: followingKuhn-Tucker necessaryconditions:
PaY0 + A(Pa8o - Ga) + A(Paas0
-

Gaa) =

0,

< ? PbY + A(Pb8o-Gb)+ (-Gb) L(Pab8?-Gab)+ where y0 = (m, - w ) - (7rF - w?) and 80 = v - v?. Since a0 > 0 maximizes the

agent's expected utility, eo also satisfies


Paso= Gas

The first and the last of the three displayedequationsyield


Y =p
Pa

(Paaao - Gaa) = p IGaa- p


Pa

Pa!

Ga) > ?

Substitutingthe last two equations into the second inequalityand dividingby Ga(e0)> 0 yield condition (3). The multipliersA and A are determinedby the Kuhn-Tucker conditionsfor wo and wo as follows:
pu=p(e0)(1
-p(e?))(pa(e0))
1) +

([V?(ws)] (1 -p(e0))([Vf(wF))]

[V'(wF)] ) > O0
> 0. Q.E.D.

A =p(e0)([V'(wo)]

The next proposition,our first main result, providesa sufficientconditionfor teamworkto be optimalwhen agents'marginaldisutilitywith regardto helping effort is zero at zero helping effort. solutionis team4: SupposeGb(a0, 0) = 0. Thenthe second-best PROPOSITION workif thefollowingconditionholds:
(4) Gaa(eO) + (Pab(e) Ga(e0) + b o) Paa(ee) ) Pa(eJJ

PROOF: By Lemma 2, if the left-hand side of (3) is strictlypositive for all


positive A,,u, and 6, offering (w?,e?) to both agents cannot be the overall

second-best.This implies that offeringany other optimalindependentcontract, if any, to each agent (whetheror not both agents are offered the same contract)

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

625

is also suboptimal, and hence the second-best solution must have positive helping effort. Substituting Gb(e') = 0 and Gab(e0) = 0 (the second of which follows from Assumption 2 (ii) and the twice-continuous differentiability of G( )), and dividing by pb(e0)/pa(e0) > 0 yield
(5) A +
Gaa_eo__

Pab(e

_Paa(eO)

G(e

( pb(e )

>0
J

pa(eo)

This holds for all A > 0 and pu> 0 if condition (4) holds.

Q.E.D.

REMARK: This result does not rely on whether the second-best solution is symmetric (e, = e2) given that teamwork is optimal. It might be asymmetric such that both agents allocate more of their efforts to the same one of the tasks. Similarly, the result still holds without Assumption 3 if the extremely asymmetric case where both agents work only on one task is called teamwork as well as the case of mutual help. Such asymmetric solutions might arise when Pab is very high. It is not easy to obtain the exact conditions for the optimal teamwork contract to be symmetric because of the nonconvexity of the implementation cost function C( ).13 Note that by Proposition 3 when the marginal disutility of helping effort is zero at zero help, the first-best contract always induces positive helping effort (given a** > 0). Proposition 4 shows, however, that in the second-best situation, the principal may prefer zero helping effort unless (4) holds: There may be a case where the first-best solution is teamwork while the second-best is an independent contract. Condition (4) represents the effects of introducing "a small amount of teamwork" on the principal via the incentive compatibility constraints. To clarify the condition, let (a, () be the optimal response functions of the agents, that is,

(a(e;w),fl(e;w))

argmaxU(w,a',b',e)
(a', b)

where w is monotone increasing. Note that a(-) and fl8() are single-valued because of the strict concavity of U(w, *, e). At the optimal independent contract (w?, eo), these response functions satisfy the following first-order conditions: (6) Un a,? pae(wo,
(w0, a,f, Ubn

) = pa(a, 0) [ V(w?) - V(w?)] -Ga(a, 0) =0, eo) = -Gb(a, ) < 0.

When Gb(a 0) = 0, increasing agent n's helping effort marginally from zero does not affect his own effort level. Then the only effect is on agent k's own
13 I am gratefulto MartinHellwig for drawingmy attention to the issue of asymmetric solutions.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

626

HIDESHI ITOH

effort level. This effect can be obtainedfrom the first equationin (6) as follows:
(7)

Db (eOwO)

Gaa -Paa[V
Pab/Pa Gaa/Ga

v]

-Paa/Pa

By (7) we obtain the followingcorollaryof Proposition4.


COROLLARY 1: Suppose Gb(a0, 0) = 0. Then the second-best solution has positive helping effort if

(8)

Pb(a

O) +Pa(ao,o)

OaO(0,;w)

>0.

If condition(8) holds, the marginalincrease in help inducedby the introduction of a small amount of teamworkmakes the incentive compatibilityconstraintsless stringent,and hence the principalis better off. The firstterm of the left-hand side of (8) is the direct effect of marginal increases in agent k's of task n. The second term represents helping effort on the success probability the indirect effect on the success probabilitythrough the external effect of marginalincreasesin agent k's help on agent n's own effort level. The former directeffect is alwayspositivesince helpingeffortis productive.The latter effect is positive, in particular,if Pab> 0, which means that the marginalproductivity of own effort is increasingin helping effort. We call this a complementarity
condition. Equation (7) shows that (when Gb(e0) = 0) the complementarity

condition is equivalent to the condition that, other effort levels fixed, when This agent k increaseshis helpingeffort, agent n adjustshis own effort upward. condition holds, for example, when the success probabilityis multiplicatively
separable in a and b; p(a, b)
=

q(a)r(b) with q( ) and r(-) increasing and

concave. conditionis far from necessary.Teamworkis Of course the complementarity clearly optimal when success probability is additively separable, that is, p(a, b) = q(a) + r(b) with q( ) and r(-) as above, so that there is no externality between two agents' incentive problems. This is simply because the helping effort is productiveand inducinga small amount of help is costless. This also implies that if there were only a single agent, the principalwould alwaysinduce him to work on anothertask when his marginaldisutilityof effort into the new task is zero. A more interestingresult is that even if an increase in an agent's helping efforthas a negativeexternality on the other agent'sown effortlevel (the case of condition(8) can be satisfied.For example,suppose that own effort free-riding), and helping effort are perfect substitutes;p(a, b) = q(a + b) with q(*) increasing and concave. Then because of the diminishingreturnsto total effort, each agent reduces his own effort level when the other agent increases his help.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

627

However,this negative effect is not large enough. Since Pa =Pb and Paa Pbb, the only net effect in (4) is through Gaa/Ga. This term is positive so that dac/db > -1: an increasein helpingeffort does not reduce the own effortby the same amountbecause the decrease in own effort reduces the marginaldisutility
of own effort: teamwork can reduce the cost of inducing each agent to work on his own task. Similarly,if Pab < 0 holds but Paa is negative enough, the principal

teamworkbecause the highermarginalproducmaybe better off by introducing tivityof own effort reduces the incentivecosts. The interpretationof the sufficientcondition is so far built on the effect of constraints.Suppose instead (4) holds teamworkon the incentivecompatibility with equality.This impliesthat the marginalchange of the optimalindependent contracttowardteamworkdoes not affect the tightnessof the incentivecompatibility constraints. The principal is then better off since the participation constraints become less stringent: a small amount of help induced by the marginalchange of the independentscheme increases the success probability while it does not increasethe disutilityof effort. The sufficientcondition for teamworkto be optimal fails to hold only if an of own effort (Pab < 0) SO increase in help decreases the marginalproductivity much that this negativeexternaleffect (a high degree of free-riding)dominates
the other positive external effects through Paa/Pa and Gaa/Ga.

It is well knownthat the optimalwage schedulein the standardhidden action model can be interpretedin statisticalterms.(See Hart and Holmstrom(1987).) Our result can also be interpretedfrom the perspectiveof statisticalinference.
pa(e0)

The sufficient condition (4) holds if Pab(e0)/Pb(e0) > paa(e0)/pa(e0). When this is equivalent to the statement that helping effort affects the =pb(e0),

inference on agents' own effort choice more effectivelythan own efforts do.14 For example, if Pab> 0, increases in helping effort make the detection of agents'own effort easier. Since Paa< 0, however,increasesin own effort make the detection more difficult.Even if Pab< 0 the condition above implies that the negativeeffect of helping effort on the inferenceof own effort is at least as of own small as that of own effort. Thus, as long as the marginalproductivity effort and helping effort is the same, the better statisticalpropertyof helping effort is sufficientfor teamworkto be optimal.
5. A NONCONVEXITY IN THE OPTIMALTASK STRUCTURE

In the previoussection, we have obtaineda sufficientconditionfor teamwork to be optimalunder the assumptionthat the marginaldisutilityof helpingeffort is zero at zero help. What if it is strictly positive as in the case where the disutility term only depends on the total amount of effort? The result is different.It turnsout that the optimalindependentcontract(w?,e?) drastically alwayssatisfiesthe necessaryconditionin Lemma2. In the proof of the lemma,
14

If Pa =Pb

at eo, the inequality above is equivalent to d(pa/p)/db > d(pa/p)/da


<d(Pa/(A

and

d(-Pa/(l

-p))/db

-p))/da.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

628

HIDESHI ITOH

conditionsfor wo? the multipliersA and ,u are determinedby the Kuhn-Tucker and w?. However,6 cannotbe specifiedfromthe conditions.Thus,if Gb(e0) > O, by taking 6 sufficientlylarge, condition (3) is always satisfied at (w?, eo). In other words, there alwaysexists 6 such that (3) holds at (w?, e?); the left-hand side cannot be strictlypositive for all 6 > 0. An implicationof this negativeresult is that marginalchangesof the optimal independent contract toward teamworkcannot induce positive helping effort: By Proposition2, such a marginalchange is valuableonly if it induces an agent to exert positive helping effort, and it in fact does when Gb(a0,0) = 0. When Gb(a0, 0) > 0, however,since the strict inequalityholds in the second equation in (6), the optimal response f3(e0;wo) of an agent does not respond to any marginalchange from wo or eo. For example, consider a small change of the
= wo? + Evjn for positive E and independent scheme of agent n, given by Wc!'(E) < > 0 such as E 1J" we can affect his marginal nist and i,j &. By choosing 'F 0, effort when E is small: with to his helping regard expected utility bn
Ub el o=pb(wS -WF)

[ pV (Ws)

+ (1 -P)V

'(w)]

> 0

is the cross partialderivativewith regardto bnand e. However,the where Ubn., small rewardfrom choosing a small amountof positive helping effort is always dominatedby the resultingincrease in disutility. The principal hence cannot induce a risk averse agent to provide a small amount of help unless his wage substantiallydepends on the outcome of the other agent'stask. Thus, we cannot obtain sufficientconditionsfor teamworkto be optimal by simply examininglocal conditions at the optimal independent contract.We can howeverobtain a general, strongerresult that small amounts
of helping effort are always suboptimal from the principal's point of view. This is

because a small amountof help increasesthe principal'srevenue infinitesimally higherincentivecosts for the while its implementation requiresdiscontinuously zero help. principalthan the costs of implementing 0)> 0, then C(e1, e2), the mini5: Fix an> 0 and ek. If Ggn(an, PROPOSITION mum expected cost to the principal of implementing (el, e2), is discontinuous at bn =O as bn increases.
PROOF:Let n = 1. We fix e2 and a, > 0, and let b, = E > 0 go to zero. For each E > 0, we have a unique optimal wage schedule W[n(E) that implements (e1(E),e2) at least costs where e1(E)= (a1, e). (The optimal scheme is unique

because the ImplementationProblem has a strictlyconvex objective function with linear constraints.)These wages WnJ(E) satisfy (PC) and (LNIC) so that U = 0. Let vOJ(E) = Vn(w1j(EJ)). Then by Lemma 1, V V?(J)> Vn(E) and v/n(8) ( VF (E) hold for i, j = S, F, with strict inequalities for n = 1. Looking along a subsequenceif necessary,we can assumethat each w/j(E) convergesas E -O 0, to E Yt. We claim that (wj) implement(e1(O),e2); fixing e2, if e1(O)is not some w11 agent l's best response at wages wh., then there is some strictlybetter response

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

629

than e1(E) for el, which would then be strictly better response to e2 at w2-(W) some (sufficiently small) E > 0, a contradiction. Similarly, we can show that e2 must be a best response to e1(0) at wc2j. Now if C() is continuous at E = 0, (WCJ)will be the optimal wage scheme for (e1(O),e2); by the continuity of wages and responses, the principal's net profits are continuous as E -* 0. However, by continuity VCJ= Vj(W satisfy
VS > vF

and

vC > VC

for i, j = S, F,

and also satisfy (LNIC) with Ul = 0, that is,


(9) Ubl(w, e1(0), e2) =pb(a2,0)[pl(al,
b2)Sl +

(1 -pl(al,

b2))F1

-Gl(al,0)

=0

where a' = Vi

- Vl. By (9) and Gl(a1, 0) > 0, either 81 > 0 or 81> 0 must hold: the optimal scheme of agent 1 for (e1(0), e2) must be contingent on the outcome of task 2. This contradicts Proposition 2. Q.E.D.

-V

and 81= vl

Figure 1 shows a graph of the implementation costs as a function of b1 = 8. When b1 > 0, however small, the principal requires imposing risk on agent 1 through task 2 and this has the first-order effect under the assumption Gb(a1, 0)> 0. However, in the limit b1 = 0, there is no need for such risk sharing. COROLLARY 2: Fix an > 0 and ek and suppose Gn(a, 0) > 0. Then for sufficiently small 8 > 0, the principal's expected net profits at bn = e are strictly smaller than those at bn = 0. PROOF:Let n = 1. By Proposition 5, C(a1, 8, e2) is discontinuously larger than C(a1, 0, e2). On the other hand, R(e1, e2) is continuous at b1 = 0 as b1 increases. Thus, for sufficiently small E > 0, R(a1, 1, e2) -C(a1, E, e2) < R(a1, 0, e2) Q.E.D. C(a1, 0, e2) holds.
C (el (*), e2)

lmC(e ( l bl),e2)
44

FIUR

C(el(O),e 2)

0-

bi
FIGURE 1

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

630

HIDESHI ITOH

By Corollary2, when an agent's marginaldisutilityof zero help is strictly positive,the principalwants him to either specializein his own task or provide substantialhelp with the other agent. This result may partiallyexplainwhy we team-basedwage observe either strict individualwage schemes or significantly schemes: intermediate cases, where wage schedules of workers depend on in a very limited manner,are rarelyobserved. performances co-workers' Note that this result has nothingto do with multi-agentissues. The resultcan be applied in situations where the principal determines whether her agent should work on a single task or performmultipletasks. In fact, a similarresult holds even for the standardone-agent model where the agent's effort is chosen from a closed interval[a, ii]. That is, if C(a) is the least cost of implementing action a and if the agent is risk averse (and other technical conditionson the distributionof outcomes and the disutilityof effort are met), C( ) increases discontinuouslyat a. Note in this regard that Grossman and Hart (1983, Proposition 1) show that C(-) will, in general, be lower semicontinuous.We extend this, then, to give conditionswhere it is definitelydiscontinuousat a. in the As a final remark,we note that the result is similarto the nonconcavity value of informationin Radner and Stiglitz (1984). They show that a small amountof informationhas a negativemarginalnet value (as a small amountof help for the principalin our setting)wheneverthe marginalcost of information is strictlypositive.The differenceis that in their model, there is no discontinuity because their model is not in the informationcost function at zero information concerned with incentive problems. The discontinuityarises in our model because of the existence of the incentiveproblemdue to hidden action. We close this section with a conjecture.When the marginaldisutilityis always positive,the first-bestsolution may have zero helping effort, dependingonly on the comparison of the marginal productivityand the marginal disutility of helping effortwith those of own effort.Thus, if the incentiveeffect discussedin the previoussection is sufficientlyfavorableto teamwork(e.g., a high level of there may exist a case in which the first-bestsolution is an complementarity), independentcontractwhile the second-bestsolution is teamwork.
6. CONCLUDING REMARKS

The main objective of this paper was to find what determineswhether to induce teamworkor unambiguousdivision of labor. We found that there are two importantfactors in determiningthe optimal task structurefrom incentive points of view: strategicinteractionbetween agents and their attitudes toward multipletasks.Althoughmanymodernproductionactivitiesinvolve performing interdependenceamongand multipletasks performedby workers,the resultsof the paper suggest that the incentive effects of introducingthese features be carefullyevaluated. If an agent's marginaldisutilityof performingan additionaltask is zero, we were able to obtain a sufficientcondition for teamworkto be optimal. Teamwork is optimalif each agent increaseshis own effort respondingto an increase in help from the other agent. Teamworkcan also be optimal without such

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

631

complementarity. Very often people free ride on others' help and reduce their own effort. In such a case, more factors affect the task design problem. Suppose that each task is so monotonous and boring that the worker's marginal productivity on that task declines or his marginal disutility increases drastically as he works harder. If this positive effect of free riding dominates the negative effect, teamwork is still optimal because it reduces the cost of inducing each worker to work on his own task. This result corresponds to the assertion by behavioral scientists that job enlargement and enrichment can motivate workers to work hard. On the other hand, the situation becomes more complicated when agents are reluctant to provide even small amounts of help. For example, complementarity is no longer sufficient for teamwork to be optimal. We have to consider costs of inducing an agent to perform multiple tasks, which are substantially larger than the incentive costs for the specialized agent. The result is the nonconvexity of the optimal task structure: The principal wants either a specialized structure or a substantial teamwork. Though not discussed in the main text, there is another important problem associated with teamwork; the problem of collusion among agents. As Mookherjee (1984) points out, there are two kinds of collusion problems. The first one comes from the multiplicity of Nash equilibria: Given wage schedules, there may exist another Nash equilibrium preferred by both agents to the one the principal wants to implement. This is a serious problem which arises only when teamwork is introduced in our model.'5 The other collusion problem is that the agents may collude to choose some cooperative effort pair that is different from the Nash equilibrium pair the principal attempts to implement. Introducing teamwork enables the agents to have full information about each other's actions, and hence is likely to promote such collusion.16 As an extension of our model, we briefly mention the case of more than two agents. It is straightforward to extend our two-agent model to the case of N agents: Agent n chooses, instead of a two-dimensional effort variable, an N-dimensional effort vector (ejn', en), where en is his own effort level and ek (k 0 n) is his helping effort level for agent k. The success probability of task n is a function of the inputs to that task, (en , en). The main results in this paper can be extended to this case. One important difference appears in the analysis of the sufficient condition for teamwork. When N > 3, whether or not agent n and agent k form a team generally depends on the effort choice by agent 1 (1 0 n, k) as well: a "good" relation between agent n and agent k (e.g. complementarity between them) may be destroyed when agent 1 joins their team. Thus the sufficient condition for teamwork to be optimal is not as simple and clear as the conditions in Proposition 4 and Corollary 1. If we assume that
15 One approach to this problem, due to Ma (1988), is to enlarge the strategy sets of the agents so that the principal can implement what she prefers as a unique perfect Nash equilibrium of the effort choice game by agents. 16 However, this may not be a disadvantage of teamwork. Itoh (1990) shows that the principal, by designing incentive schemes appropriately, can be better off when agents collude. See also Holmstrom and Milgrom (1990) and Ramakrishnan and Thakor (1989).

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

632

HIDESHI ITOH

the relation between two agents is independent of the third agent (which I think is not an unreasonable assumption), our sufficient conditions are still valid: The comparison between two agents is sufficient to determine whether they become members of the same team to help one another. This independence assumption therefore has the following interesting implication on the optimal team size. Suppose that all the agents are identical. Then if they have zero marginal disutility of helping effort at zero help, there is no particular cost of increasing the size of the team. The principal chooses either N-size teamwork or no team at all. However, if the agents care only about the total amount of effort so that their marginal disutility of help is not zero at zero help, increasing the size of the team raises the agent's marginal disutility of help, so that there may exist a limit on the team size.'7 In this paper, we focused on the incentive aspect of teamwork, following the literature on the moral hazard in the principal-agent relationship. A shortcoming of this approach is that we cannot obtain many clear predictions on real management policies.18 In particular, in our model, agents' abilities, characteristics, and task characteristics are common knowledge among all the relevant parties. Also we assumed away all dynamic aspects of task design problems. In more realistic settings incorporating learning, reputation, information asymmetries, or firm-specific human capital accumulation, we will be able to pursue our further understanding of teamwork and task structures. Department of Economics, Kyoto UniversitySakyo-ku, Kyoto 606, Japan
Manuscriptreceived December 1988; final revision receivedApril, 1990.

APPENDIX In this appendix, we obtain sufficient conditions for the first-order approach to be valid. The procedure of the proof is similar to that of Rogerson (1985). We first show that the wage schedule in the solution of (RP) is monotone increasing: given an outcome of one task, the wage under success of the other task is at least as high as the wage under failure. Then we prove that the expected utility of each agent is concave in his effort variables under the condition given in Assumption 5. We introduce another relaxed problem, denoted by (RP + ), in which the principal chooses (w , w2) and (el, e2) to maximize the objective function in problem (OP) subject to the participation constraints (PC) and the following local Nash incentive constraints (LNIC +): (LNIC + )
Ua,(Wn,enf,ek) =0

forn= 1,2;
bn *Ub(Wn enf,ek)=O

Ub(w, en, ek) 6 0 and

forn=1,2.

The ct. traints (LNIC + ) are the standard local conditions for (NIC).
17Our model might also offer an explanation of increasing returns to size from the incentive viewpoint. Managers of two separate firms cannot share their tasks, and hence are contracted independently. When the firms are merged, under some conditions, the owner would like to design new interdependent contracts in order to induce them to share their tasks, and thereby increasing returns to firm size might arise. I am grateful to George Mailath for suggesting this. 18 In earlier versions of the paper, I obtained some properties of the optimal teamwork contract.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

633

For each n, let An be the Lagrange multiplier for (PC), An and (n be the Lagrange multipliers for the conditions for an and bn in (LNIC) in the relaxed problem (RP), respectively. The Kuhn-Tucker necessary conditions for the optimal wage w/j is obtained as follows:

(Al)

dPP (enl,ek)/danl

liP'2(n,e)l

nk)/n

for i, i

Lemma A-1 shows that the wage schedule for each agent is monotone increasing.
LEMMA A-1: Let (w, w2) be the optimal wage schedule of (RP) for (el, e2) such as an >0, n = 1,2. Then w ' > wn and wni > w[n for i, j =S, F. In addition, the equality holds if and only if
bn = 0-

PROOF:For i, j = S, F, (A2) dPslda

pn

dPjF/dan
P

=-

-pa p

pFnj AP / n pn jln

Pb

1 pkn p

, pniSab

pk(n_V)+

-p)V

F-F)

Ga6

If (l < 0, then by (Al) and (A2), wsn S wF1 must hold for each j =5S,F. Then

bn>0 ?
i=S,F.

for all an > 0, which is a contradiction. Hence Mn > 0. Then w5nj> if wFjfor all j =5S,F. Similarly, must hold, so that w/i5 > w/i for i =S, F. If bn=0O,then by (Al), w/j5= w,'4 for n
Q.ED.1F

The next lemmashowsthat the solutionto (RP) is in the constraintset of (RP + ).


un = Pnk[kP n Vsns5 ) + ( 1- _k) (V F-VF )n Ga]-G an< 0 LEMMA A-2: Suppose that(eA, (w1 w2) and e2) such as an > 0 for n = 1,T2 solve (P). satisfy (LNIC +?). PROOF: Clearly it is sufficient to show F Ub7

Thenthey

0. If bn> 0, by (LNIC), U,7 =0. Thus, (LNIC +) is

satisfied.Next, supposebn= 0 and Ub, > 0. That is,

Thus, either v~55>VSF or v5> VF> must hold. Then by Lemma A-i, > for allim=iS, F. vj>W vTe5 Therefore bn > 0, which is a contradiction. Thus, when bn= 0, Uj 6 0 must hold, so that (LNIC +)
issatisfied. Q.E.D.
LEMMA A-3: (w1u w2) and (el, e2) such as an >

0 for n = 1,2 solve (RP) if and only if they solve

(RP+).

PROOF:The "only ifs' part is trivial by Lemma A-2. Concerning the "i' part, suppose (wL , w2) and (et , e2) solve (RP + ) while they do not solve (RP). (is, and (e, e2) solve (RP). Denote ,2) Let by UP(-) the principal's expected net profits. Then since (w, w2) and (e, e2) satisfy (LNIC), the

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

634

HIDESHI

ITOH

optimality of (w, w^2) and (l, e2) results in


up(w1,
2,ele2 > up(w1, 2,ele)

However, by Lemma A-2, (wii, w2) and ( Y, e) satisfy (LNIC + ), so that they are in the constraint Q.E.D. set of (RP + ). A contradiction. Based on Lemma A-1, Lemma A-4 provides sufficient conditions for the first-order approach to be valid. LEMMA A-4: Suppose that (w1, W2) solve (RP) for (el, e2) such as an > 0, n = 1, 2. Thenfor each n, the expected utility of agent n is strictly concave in en if either one of the following holds: (i) is concavein en and h' is concave;(ii) PFJ = (1 _pn)(1 _ k) is convexin en and hn is PSS =pnpk convex.
PROOF:

Suppose that condition (i) holds. The expected utility of agent n at w' is given by
Un(wn,.)=VF
+pn(.)(VnFVnF) +p n()pk(.)[(Vn +p(.)(VnS-VnF) -VnF) -(VnS-VFn)]

(A3)

-Gn(

= pnpk When P~Sn

LemmaA-1, VnF-

is concave in >0 and VFF > 0 and

en, clearlypn is concavein an and pk is concave in bn. And by 0. ih n - Vn iphner is concavein en, F VFF > O. Thus, if the term with SS =pnpk

Un is strictly concave (because Gn is strictly convex). To show this, it is sufficient to show n - VnF > Vn - VFn. By (Al) and (A2), we obtain (A4)
(vFn) ) -hn hn( = hn( n )-hF n(FF) = npb nbnPb( + 1_pk)

If the optimal solution has no positive helping effort by agent n, the right-hand side of (A4) is zero. Otherwise, the right-hand side is strictly positive. Then, since hn is an increasing function when agent n is risk averse, it follows from (A4) and the hypothesis that hn is concave that VnS - VnF >
vF -vFF.

To show that condition (ii) is sufficient, let qm = 1 -pm and express the expected utility of agent n by using qfn and qk. Then follow the steps similar to the case of condition (i). Q.E.D.
LEMMA

and (el,

e2)

A-5: Suppose either one of the conditions (i) and (ii) in Lemma A-4 holds. Then (w1, W2) such as an > 0 for n = 1, 2 solve (RP) if and only if they solve the original problem (OP).

PROOF: The "only if' part is by Lemma A-4. The proof of the "if' part is similar to that of Lemma A-3. Q.E.D.

The sufficiency of the condition (i) or (ii) in Lemma A-4 holds in the case of more than two outcomes. Suppose that for n = 1, 2, there are Mn possible outcomes 1, , Mn of task n, and for i= 1, Mn, let 7rn be the revenue from task n when the outcome of task n is i with 17-n < ... Also for i = 1,, <1TM. Mn and j 1,, Mk, let FiJ(en, ek) be the probability that, given effort levels (en, ek), the revenues from task n and task k are at least as high as .n7[ and 74, respectively. Suppose that condition (i) holds. The generalization of the condition on the joint probability in (i) is that for n = 1, 2, FI is concave in en for all i and j. The expected utility of agent n at wn is given by
Mn Mk

un(wn,

) = 1=1 j=1

[(v

n-

(Vn

n-V/1

Vn

j1)]FJ(.)

-Gn(

where VnO = Vn = 0 for all i and j. Since the generalization of Lemma A-1 is straightforward (with MLRP) and condition (i) in Lemma A-4 results in the expression in the brackets nonnegative for all i and j, the expected utility is concave. We have not proved that (n > 0 when the optimal helping effort level of agent n is zero. To show this, we follow Rogerson (1985) and define the doubly relaxed problem, denoted by (DRP), which is

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

MULTI-AGENT

SITUATIONS

635

the problem (RP) with the following constraints (DLNIC) in place of (LNIC): (DLNIC)
U,(w',ef,ek)=O
bn Ubn(w

forn=1,2;

,enl, ek) > 0

for n = 1,2.

Note that the second line is not an equation. Then the Kuhn-Tucker necessary conditions for the solution to (DRP) immediately lead to the multiplier (n for the second inequality in (DLNIC) is always nonnegative. What must be proved is that the solution to (DRP) satisfies (LNIC).
LEMMA A-6: Suppose that (w1, w2) and (el, e2) such as an they satisfy (LNIC).
> 0

for n = 1, 2 solve (DRP). Then

PROOF: We have to show that (w , w2) and (el, e2) satisfy the second condition in (DLNIC) with equality. If (n > 0, this follows immediately from the complementarity slackness. Now suppose (n = 0. Then the proof similar to that of Lemma A-1 shows that the wage schedule of agent n does not depend on the outcome of task k. Thus, if bn > 0,
bn- Ubn = bn(-Gn)<0.

Therefore bn = 0 must hold, which results in the second condition in (DLNIC) holding with strict Q.E.D. equality. By Lemma A-6, we can show that the set of the solutions to (DRP) coincides with the set of the solutions to (RP) by the proof similar to that of Lemma A-3. The multiplier (n in the Kuhn-Tucker conditions of (RP) is hence nonnegative.

REFERENCES ALCHIAN, A. A., AND H. DEMSETZ (1972): "Production, Information Costs, and Economic Organization," American Economic Review, 62, 777-795. AoKi, M. (1988): Information, Incentives, and Bargaining in the Japanese Economy. Cambridge: Cambridge Univ Press. BAIMAN, S., AND J. S. DEMSKI (1980): "Economically Optimal Performance Evaluation and Control Systems," Journal of Accounting Research, 18, 184-220. DERTOUZOS, M. L., R. K. LESTER, R. W. SOLOW, AND THE MIT COMMISSIONON INDUSTRIAL PRODUCTIVITY (1989): Made in America: Regaining the Productive Edge. Cambridge: The MIT Press. DRAGO,R., AND G. K. TURNBULL (1987): "Competitive and Non-Competitive Incentives in Team Settings: Notes toward a Theory of Promotion Systems," mimeo. (1988): "Individual versus Group Piece Rates under Team Technologies," Journal of the Japanese and International Economies, 2, 1-10. GREEN, J. R., AND N. L. STOKEY (1983): "A Comparison of Tournaments and Contracts," Journal of Political Economy, 91, 349-364. GROSSMAN, S. J., AND 0. D. HART (1983): "An Analysis of the Principal-Agent Problem," Econometrica, 51, 7-45. HACKMAN, J. R., AND G. R. OLDHAM (1980): Work Redesign. Reading: Addison-Wesley. HART, O., AND B. HOLMSTR6M (1987): "The Theory of Contracts," in T. F. Bewley, Advances in Economic Theory, Fifth World Congress. Cambridge: Cambridge University Press, pp. 71-155. HOLMSTR6M, B. (1982): "Moral Hazard in Teams," Bell Journal of Economics, 13, 324-340. HOLMSTR6M, B., AND P. MILGROM (1987): "Aggregation and Linearity in the Provision of Intertemporal Incentives," Econometrica, 55, 303-328. (1989): "Multi-Task Principal-Agent Problems," mimeo, Yale University. (1990): "Regulating Trade among Agents," mimeo, Yale University. ITOH,H. (1990): "Collusion, Incentives, and Risk Sharing," mimeo, University of California, San Diego.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

636
KAGONO,

HIDESHIITOH

T., I. NONAKA,K. SAKAKIBARA,AND A. OKUMURA (1985): Strategic vs. Evolutionary Management: A U.S.-Japan Comparison of Strategy and Organization. Amsterdam: NorthHolland. LAZEAR, E. P. (1989): "Pay Equalityand IndustrialPolitics," Journalof PoliticalEconomy,97, 561-580. LAZEAR, E. P., AND S. ROSEN (1981): "Rank-Order Tournaments as Optimal Labor Contracts," Journal of Political Economy, 89, 841-864. LINCOLN, J. R., AND K. McBRIDE (1987): "Japanese Industrial Organization in Comparative Perspective," Annual Review of Sociology, 13, 289-312. MA, C. (1988): "Unique Implementation of Incentive Contracts with Many Agents," Review of Economic Studies, 55, 555-571. McAFEE, R. P., AND J. McMILLAN (1986): "Optimal Contracts for Teams," mimeo, University of Western Ontario. MILGROM, P. R. (1981): "Good News and Bad News: Representation Theorems and Applications," Bell Journal of Economics, 12, 380-391. MIRRLEES, J. (1975): "The Theory of Moral Hazard and Unobservable Behavior-Part I," mimeo, Nuffeld College, Oxford.
MOOKHERJEE,D.

(1984): "OptimalIncentiveSchemes with Many Agents," Reviewof Economic J. E.


STIGLITZ(1983):"Pricesand Incentives: Towarda

Studies, 51, 433-446.


NALEBUFF, B. J., AND

GeneralTheoryof

Compensation and Competition," Bell Journal of Economics, 14, 21-43. RADNER, R., AND J. E. STIGLITZ(1984): "A Nonconcavity in the Value of Information," in M. Boyer and R. E. Kihlstrom, Bayesian Models in Economics. Amsterdam: North-Holland, 33-52. in Agency: versusCompetition RAMAKRISHNAN,R. T. S., AND A. V. THAKOR (1989):"Cooperation Incentive Problems, Diversification, and Corporate Mergers," mimeo, Indiana University. Problems," Econometrica, Approachto Principal-Agent ROGERSON, W. P. (1985): "The First-Order 53, 1357-1367. WATERMANJR., R. H. (1987):The Renewal Factor: How the Best Get and Keep the CompetitiveEdge. New York: Bantam Books.

This content downloaded from 165.123.34.86 on Wed, 11 Sep 2013 08:41:16 AM All use subject to JSTOR Terms and Conditions

You might also like