Incentive Design for Hidden Actions and Noisy Outcomes

This direction is split into other directions: moral hazard. Its last version stays readable and citable here.

Version 1 of 164 papers69 conceptsfrozen 2026-10-04

In a perfect world, a manager would pay a worker based on how hard they work. But in the real world, effort is often invisible. A programmer might spend all night writing elegant code or all night playing video games; a security expert might implement rigorous protocols or simply get lucky that no one attacked. This information asymmetry—where an agent's actions are hidden from the "principal" who hires them—is known as moral hazard.

Because the principal cannot see the effort, they must instead reward the outcome. This shift from paying for input to paying for output is the core of contract theory. Recent research has moved this economic problem into the realm of computer science, treating the design of these incentives as an algorithmic challenge.

The Problem of Noisy Proxies

The fundamental difficulty in contract design is that outcomes are "noisy." As Matteo Castiglioni and Nicola Gatti note in Multi-Agent Contract Design, a project might fail despite high effort due to bad luck, or succeed despite laziness due to good luck.

If a principal sets a reward that is too low, the agent will "shirk" to save their own costs (like time or battery power in IoT devices, as explored by Vijay Gupta in A Reputation-Based Contract for Repeated Crowdsensing). If the reward is tied too strictly to success, the agent may refuse the contract entirely because the risk of failure—and thus zero pay—is too high.

Why Simple Contracts Win

Economic theory often suggests that the "optimal" contract should be a complex, highly specific mathematical function. However, Guru Guruganesh, Jon Schneider, and Joshua R. Wang demonstrate in Contracts under Moral Hazard and Adverse Selection that finding this perfect contract is computationally "APX-hard."

Instead, they find that linear contracts—simple percentage-based commissions—are remarkably robust. While a complex contract might be theoretically better, a linear contract provides a strong approximation of the maximum possible welfare and is much easier to implement. This explains why we see commissions and profit-sharing everywhere in the real world: they are "good enough" solutions to an otherwise intractable optimization problem.

Scaling to Teams and Learning

The challenge intensifies when a principal hires a team. In Fair Team Contracts, Junjie Chen and Yingkai Li explain that individual contributions are often obscured by the collective result. This creates a "free-rider" problem where agents might under-invest effort, hoping their teammates will carry the load.

To solve this without knowing the agents' internal costs upfront, researchers are turning to machine learning. Michael I. Jordan and colleagues in The Sample Complexity of Online Contract Design frame this as an online learning problem. The principal "learns" the optimal contract over multiple rounds of interaction, balancing the need to explore different payment structures with the need to exploit the best-known incentive to keep agents productive.

Automating Design with Deep Learning

Traditional contract design relies on "pen-and-paper" models that struggle with high-dimensional data. To bridge this gap, Tonghan Wang and Paul Dütting introduced Deep Contract Design via Discontinuous Networks.

They developed a specialized neural architecture called Discontinuous ReLU (DeLU). Standard neural networks are continuous, but contracts often have "jumps"—for example, a bonus that only kicks in after a certain threshold is met. By explicitly modeling these discontinuities, DeLU can automate the creation of optimal contracts for complex scenarios, such as managing systemic cyber risk in enterprise networks (Juntao Chen and Quanyan Zhu, Dynamic Contract Design), where the relationship between effort and security is highly nonlinear.

The Path Forward

As AI agents begin to act as autonomous workers in labor markets, understanding these incentives becomes critical. As argued by Simpson Zhang and Mihaela van der Schaar in Agents Require Metacognitive and Strategic Reasoning, these agents must be designed to navigate the same economic forces—moral hazard and adverse selection—that humans have faced for centuries. The goal of this research direction is to ensure that even when we cannot see what an agent is doing, we can structure the world so that their best interest remains aligned with our own.

Papers in this version

Showing 60 of 64 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_incentive_design_for_hidden_actions_and__12ef9983,
  title = {Incentive Design for Hidden Actions and Noisy Outcomes},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/incentive-design-for-hidden-actions-and-noisy-outcomes/rdr_12ef99831ad74af9a5cfd88af1250a24},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_incentive_design_for_hidden_actions_and__12ef9983_v1,
  title = {Incentive Design for Hidden Actions and Noisy Outcomes},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/incentive-design-for-hidden-actions-and-noisy-outcomes/rdr_12ef99831ad74af9a5cfd88af1250a24/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}