Simple output-sharing contracts can remain optimal when employers learn

An employer rarely knows every method a specialist could use, how costly each method is, or what results it might produce. Designing incentives around an incomplete picture creates a problem: a contract that looks attractive under known options could perform badly when an unfamiliar option becomes available.

In the preprint Robust Contracts with Exploration, Chang Liu asks whether learning through successive hires changes the case for simple contracts. The paper’s mathematical results show that a fixed share of output can remain optimal under a worst-case criterion, even when today’s contract helps reveal tomorrow’s possibilities. This is a theoretical finding, not evidence that such contracts outperform alternatives in actual workplaces.

Contracting without a complete map

The model follows one principal—the party offering contracts—and a sequence of agents, each working for one period. An action combines a cost to the agent with a probability distribution over output. The principal knows some available actions, including at least one whose expected output exceeds its cost, but has no probability assessment of what other actions exist.

Instead of maximizing returns averaged over beliefs about those unknown options, the principal maximizes a worst-case expected payoff guarantee. This includes current returns and weighted future returns. The worst case is evaluated again after each history, using the possibilities consistent with what has been learned. It is not a guaranteed minimum for every realized output.

A linear contract pays the agent a fixed fraction of output, with no fixed payment. The fraction can change between periods as information accumulates.

Crucially, the principal observes the chosen action, including its cost and output distribution, before contracting with the next agent. Although observable, the action cannot itself be specified as an enforceable contractual requirement. Payments instead depend on output.

Three different meanings of learning

Liu compares three rules connecting agents’ available actions.

Under independent technology, earlier choices reveal nothing about later agents’ options. Every agent has the initially known actions, but additional possibilities need not carry over. There is consequently no useful learning, and repeating the optimal one-period linear contract maximizes the overall guarantee.

Under advancing technology, available actions can accumulate but cannot disappear. Once an action is observed, the principal knows that later agents can use it. However, a previous agent’s failure to choose an unobserved action cannot rule it out later: it might become available only afterward.

Here, the paper proves that a linear contract is optimal after every feasible history, including histories involving nonlinear contracts. The proof works backward from the final period, showing that any current contract can be replaced by a linear one with at least as high a guarantee, accounting for both immediate returns and future knowledge.

Under constant technology, all agents share the same unchanging action set. Learning becomes richer. A chosen action establishes what exists; a rational agent’s choice also excludes any action that would have given that agent a strictly higher expected net payoff—expected payment minus action cost—under the offered contract.

Those exclusions preserve information about previous contract shapes. After earlier nonlinear contracts, an optimal later contract can therefore be nonlinear too. In the final period, the optimal guarantee can be obtained by repeating an earlier contract, modifying one by sharing residual output, or offering an appropriate linear contract.

Nevertheless, starting from the beginning, an optimal policy uses linear contracts throughout. More generally, linear contracts remain optimal after any history consisting entirely of linear contracts. That is narrower than optimality after every possible history.

Learning can move incentives in opposite directions

A two-period calculation illustrates why contract simplicity does not mean an unchanged payment share. Here, the principal initially knows just one productive action with strictly positive cost and expected output exceeding that cost, besides the zero-cost option to quit.

Under advancing technology, the optimal first-period share exceeds the one-period benchmark. Stronger incentives make valuable actions harder to conceal in the worst case. Under constant technology, the optimal initial share falls below that benchmark: the principal also benefits from the restrictions revealed by the agent’s preferences.

What the result leaves open

The conclusions rely on risk-neutral parties, nonnegative payments, a finite horizon and a favorable rule for resolving agents’ indifference. Observing an action’s full distribution and cost is an explicitly strong assumption.

The analysis also restricts attention to deterministic contracting policies. Randomized policies fall outside its scope, and the paper notes that randomization can improve guarantees in related settings. Its contribution is therefore a conditional foundation for simple output sharing—not a universal recommendation for compensation design.