Multi-Agent Contracts

• Multi-agent contracts are formal mechanisms that delegate tasks, allocate incentives, and coordinate multiple agents with private preferences.
• They encompass various models like deterministic, randomized, and linear contracts to address incentive compatibility and computational tractability.
• Their applications span blockchain consensus, smart contract optimization, and multi-agent reinforcement learning, illustrating broad practical impact.

Multi-agent contracts are formal mechanisms by which a principal delegates tasks, allocates incentives, and coordinates the behavior of multiple autonomous agents, each with private preferences or actions, often under information asymmetry. The design, analysis, and implementation of such contracts underpin a wide spectrum of problems in economics, computer science, control theory, and distributed systems. Modern research has established general frameworks, powerful approximation results, equilibrium and robustness theorems, and revealed intricate trade-offs—such as between fairness and efficiency or between expressibility and computational tractability—across a variety of classical and emerging multi-agent settings.

1. Formal Models and Mechanism Classes

Multi-agent contracts generalize the classical principal-agent paradigm to scenarios with $n$ agents, each possessing a finite action set $A_i$, for $i \in N = \{1,\ldots, n\}$ and facing potentially complex interdependent outcome distributions $F_a \in \Delta(\Omega)$, with $\Omega$ the outcome space. Each agent $i$ incurs cost $c_i(a_i)$ per action, and the principal values outcomes via $r: \Omega \to \mathbb{R}$, with all parties risk-neutral (Cacciamani et al., 2024).

Mechanism classes include:

• Deterministic contracts: The principal publishes a payment rule and a target action profile $a^* \in A = \prod_{i}A_i$. Agents play a Nash equilibrium in the induced game defined by individual expected utilities. Feasibility and incentive-compatibility constraints are encoded as a linear program over possible action and payment profiles. For finite $A$ and $\Omega$, optimal deterministic contracts exist and are computable in polynomial time (Cacciamani et al., 2024).
• Randomized contracts: The principal may randomize over action profiles, offering contract pairs $(\mu, \pi)$—where $\mu$ is a distribution over $A$ and $\pi$ are outcome-conditional payment functions. Here, incentive-compatibility is respected in expectation over the recommended and deviating actions. Powerful relaxing results establish that randomized contracts may strictly outperform deterministic ones, even yielding unbounded gaps in some instances. However, the optimal value may not be attained (supremum only), requiring $\epsilon$-approximations computable via LP relaxations (Cacciamani et al., 2024).
• Linear contracts and combinatorial extensions: When the outcome space is combinatorial in joint agent effort (such as in submodular or XOS rewards), simple linear contracts—paying fixed shares of project value—can be near-optimal, and polynomial-time algorithms yield constant-factor approximations in broad settings (Duetting et al., 2022, Duetting et al., 2024, Aharoni et al., 26 Apr 2025).
• Dynamic and resource-bounded contracts: Extensions to continuous-time, sequential, or resource-constrained environments invoke PDE/BSDE or explicit multi-dimensional budget constraints, with solution concepts based on Nash equilibrium and temporal logic (Luo et al., 2017, Ye et al., 13 Jan 2026).

2. Equilibrium Concepts and Robustness

Modern work extends beyond pure Nash equilibrium (PNE) to mixed Nash, correlated, and coarse-correlated equilibria (CCE) (Dütting et al., 24 Nov 2025). The inclusion chain PNE ⊆ MNE ⊆ CE ⊆ CCE reflects increasing generality and decreasing restrictiveness; crucially, CCE capture the limiting distribution of no-regret agent dynamics.

Key theorems—specifically black-box lifting and robustness—prove that for important classes of reward functions (notably submodular and XOS), the principal's optimal utility at the "best" CCE over all contracts is within a constant factor of the utility under the best PNE and contract. Table 1 summarizes the equilibrium-robustness landscape for main function classes:

Reward Function	U^CCE / U^PNE Gap
Submodular/XOS	Θ(1) (constant)
Subadditive	Θ(poly n); gap may be Ω(√n)
Supermodular (bin)	=1
Supermodular (arbit)	Unbounded
General monotone	Unbounded

This demonstrates that, for submodular and XOS objectives, CCE-robust contract design is tractable and enjoys strong approximation guarantees, whereas for richer classes, the equilibrium gap becomes unbounded, complicating robust design (Dütting et al., 24 Nov 2025).

3. Approximation Algorithms and Computational Hardness

Polynomial-time algorithms with value or demand oracle access achieve constant-factor approximation of the principal's optimal utility (and sometimes welfare) for submodular and XOS reward functions, even as the action set scales combinatorially (Duetting et al., 2022, Duetting et al., 2024, Aharoni et al., 26 Apr 2025). These algorithms rely on:

• Subset stability: Identifying action profiles stable under dropping components;
• Doubling lemma: Relating the value under doubled contracts to guarantee lower bounds on the principal's utility at equilibria;
• Combinatorial search and scaling: Utilizing properties unique to submodularity/XOS, such as efficient scaling of marginals and critical set reductions.

Matching lower bounds highlight the limits of computation: no PTAS exists even for binary actions, and for subadditive value functions, polytime algorithms cannot surpass an $\Omega(\sqrt{n})$-approximation (Duetting et al., 2022, Duetting et al., 2024).

4. Fairness, Non-Discrimination, and Social Welfare

Multi-agent contracts can generate widely disparate agent payments even when achieving optimal principal utility; for settings where fairness or payment parity is required, explicit trade-offs emerge. Research quantifies the price of non-discrimination (PoND)—the welfare or utility loss from enforcing equal or near-equal agent payments:

• Exact non-discrimination results in a PoND scaling as $\tilde\Theta(\log n)$ in the number of agents;
• Relaxed (ratio-bounded) non-discrimination brings the price to constant as the allowed payment ratio grows (e.g., allowing payment ratios up to $n^\lambda$ interpolates smoothly between the two extremes) (Ding et al., 23 Jan 2026).

For social welfare—aggregate surplus rather than principal profit—constant gaps between optimal welfare and principal utility are established for XOS and submodular classes (188 and 5, respectively), whereas the gap can scale as $\Omega(\sqrt{n})$ for subadditive functions and is unbounded for supermodular ones (Aharoni et al., 26 Apr 2025).

5. Applications and System Implementations

Multi-agent contract frameworks are applied across:

• Blockchain and multi-agent consensus: Swarm Contracts use multi-sovereign agents in TEEs, coordinating via off-chain protocols but achieving on-chain finality by m-of-n threshold signatures for atomic transitions (NFT minting, DAOs, cross-chain actions). They optimize for gas efficiency and trust minimization (Yang, 2024).
• Automated optimization of smart contracts: Systems such as GasAgent instantiate contractual meta-protocols between synthetic agents (Seeker, Innovator, Executor, Manager), leveraging both expert-encoded and learned patterns to iteratively optimize code, validating improvements under correctness and gas-saving constraints (Zheng et al., 21 Jul 2025).
• Collaborative control and synthesis: Assume-guarantee contracts facilitate compositional synthesis and verification in multi-agent cyber-physical or dynamical systems, enabling distributed satisfaction of temporal logic specs (LTL[$\mathcal{F}$], STL) and modular, scalable verification (Dewes et al., 2024, Liu et al., 2023).
• Principal-multi-agent RL and social dilemmas: Formalized contracts providing zero-sum transfers can resolve incentive misalignment and implement socially optimal policies in Markov games or reinforcement learning environments, even under partial observability or with limited contract expressiveness (Haupt et al., 2022).

6. Open Problems and Research Directions

Active areas of research include:

• Extending positive approximation and robustness guarantees to large-scale and succinctly represented action spaces;
• Characterizing the power of randomization and richer equilibrium concepts (trembling-hand, sequential equilibria) in broader settings and Bayesian models;
• Integrating dynamic, multi-period, or resource-constrained dimensions, including life-cycle semantics and conservation laws in contract delegation frameworks;
• Tightening welfare-utility gap constants, especially in XOS and submodular settings, and adapting algorithms to richer feasibility constraints (e.g., matroids);
• Enhancing modular verification and synthesis of temporal contract systems for complex MAS (Cacciamani et al., 2024, Dütting et al., 24 Nov 2025, Ye et al., 13 Jan 2026, Ding et al., 23 Jan 2026, Duetting et al., 2024).

The field of multi-agent contracts continues to evolve, revealing deep connections between economic theory, combinatorial optimization, reinforcement learning, distributed systems, and formal methods. Rigorous characterizations, scalable algorithms, and empirical validation across domains are foundational to both practical deployment and conceptual understanding of contract-mediated multi-agent interaction.