Agentic Compositionality & Welfare Aggregation

• Agentic compositionality and welfare aggregation are defined as the structural integration of individual agent preferences into a global system, emphasizing risk attitudes and fairness.
• Quantile-based methodologies replace traditional additive valuations by using order statistics to capture agents’ risk profiles and decision biases.
• Advanced algorithmic designs and social choice principles underpin these models, guiding resource allocation and distributed optimization in multi-agent systems.

Agentic compositionality and welfare aggregation are central components in multi-agent resource allocation, contract design, economic theory, and learning systems, governing how individual agent-level preferences, objectives, or beliefs are combined to yield meaningful and tractable collective objectives. In modern research, agentic compositionality refers to structural properties of how local agent utility, risk, or epistemic value functions jointly define global system welfare, with consequences for tractability, fairness, and the design of aggregation algorithms or economic mechanisms. Welfare aggregation encompasses methodologies for combining agent utilities (or costs) under varying assumptions of comparability, participation rights, autonomy status, and task structure.

1. Quantile-Based Compositionality: From Additivity to Order Statistics

Traditional models aggregate agent preferences using additive valuations, where the value of a bundle is the sum of its constituent items' values. In contrast, the quantile-based model assigns each agent $i$ a quantile parameter $q_i \in [0,1]$, so that for bundle $S$, the agent's utility is the $q_i$-th quantile (order statistic) of $\{v_i(j): j \in S\}$ rather than the sum. This induces a rich space of agent attitudes:

• $q_i=0$ (pessimist), value is the worst item in the bundle.
• $q_i=1$ (optimist), value is the best item.
• $q_i=1/2$ (median), value is the median.

Quantile-based compositionality is strictly non-additive (outside of $q=0,1$) and often non-monotone; adding low-value items may reduce median or quantile utility. This model captures subtleties such as risk-aversion or focus on worst-case outcomes, unrepresentable in additive frameworks (Aziz et al., 25 Feb 2025).

2. Welfare Aggregation Objectives and Social Choice Foundations

The aggregation of agent utilities is governed both by the form of individual valuations and by the social welfare functional:

• Utilitarian welfare: $U(A) = \sum_{i=1}^n v_{i,q_i}(A_i)$ rewards total utility.
• Egalitarian welfare: $q_i \in [0,1]$0 targets the worst-off agent.

In social choice and control, the structure of the welfare aggregator is dictated by the assumptions about interpersonal comparability of utilities or costs, as formalized in (Shilov et al., 26 Mar 2025):

Comparability Level	Allowed Transforms	Canonical Aggregator
Ordinal (OLC)	$q_i \in [0,1]$1	Max-min (Rawlsian)
Cardinal Non-Comp.	$q_i \in [0,1]$2	Nash product
Cardinal Unit Comp.	$q_i \in [0,1]$3	Weighted sum
Full Cardinality	$q_i \in [0,1]$4	Weighted sum + inequity formal

An SCF can only aggregate information consistent with these comparability constraints, ensuring that the aggregation respects the agents' informational context and fairness axioms.

3. Algorithmic and Complexity Landscape

The computational tractability of welfare maximization varies dramatically across compositional and aggregation structures:

• Quantile Valuations: Maximizing balanced utilitarian welfare under quantile valuations is NP-hard even for the worst-item case, but admits a $q_i \in [0,1]$5-approximate greedy algorithm. For unbalanced allocations, a “scapegoat” method achieves a $q_i \in [0,1]$6-approximation. Balanced egalitarian welfare is solvable in polynomial time via bipartite matching, while unbalanced cases demonstrate sharp phase transitions: for $q_i \in [0,1]$7 in $q_i \in [0,1]$8, polynomial algorithms exist; for $q_i \in [0,1]$9 ($S$0), NP-hardness arises (Aziz et al., 25 Feb 2025).
• Multi-Agent Contracts (XOS Functions): When the project's value function $S$1 lies in the XOS (fractionally subadditive) class, there is a constant-factor gap between maximum social welfare and maximum principal utility, and polynomial-time 2-approximation algorithms are available for the welfare optimum in symmetric cases. Beyond XOS (i.e., for general subadditive or supermodular $S$2), the welfare-utility gap can be unbounded, directly mirroring the loss of agentic compositional tractability (Aharoni et al., 26 Apr 2025).
• Continuous Thiele's Rules: For continuous distributional settings, each agent is assigned a satisfaction function, and aggregation is controlled by a parameter governing the inequality aversion (IAV) of the aggregator. Smoothly tuning this parameter interpolates between utilitarian and egalitarian objectives, with explicit bounds on welfare and egalitarian loss (Wagner et al., 2024).

4. Structural Principles: Rights, Autonomy, and Coalition Composition

Advanced models, particularly in economic and post-AGI settings, explicitly track agent status (tool, delegate, agent, welfare subject), rights bundles, and institutional context. Coalition composition requires composing rights (e.g., aggregating transferable bundles or keeping track of individual autonomy vectors), and aggregating welfare via Pareto-preserving functionals (e.g., weighted sums or max-min, as justified by interpersonal comparability) (Perrier, 23 Apr 2026).

Autonomy-qualified welfare economics generalizes classical equilibrium theorems by making autonomy explicit in the commodity space and specifying seven conditions (including delegation accounting, verification, and externality closure) required for autonomy–Pareto efficiency. In the low-autonomy limit, the classical Arrow–Debreu theorem is recovered.

5. Probabilistic and Neural Agentic Compositionality

In probabilistic and learning-theoretic settings, agentic substructure is modeled by representing each subagent as a probability distribution over outcomes. The primary welfare functional is the log-score (epistemic utility), and composition occurs via weighted logarithmic pooling: $S$3. Strict unanimous welfare improvement—where every agent gains under pooling—is possible only for non-binary outcome spaces and under log-pooling, but not under linear pooling or with only two outcomes (Lee et al., 8 Sep 2025).

Key structural properties include:

• Cloning invariance and recursive composition: Pool-invariance (replacing super-agents with constituent subagents) and cloning invariance (replicating subagents) ensure stable compositional properties.
• Tilt analysis: Prevents trivial decompositions by showing that small perturbative duplications cannot generate strictly unanimous improvement unless agents are genuinely distinct.
• Alignment phenomena: Enforcing alignment on one agentic substructure (e.g., a benevolent persona in an LLM) inevitably calls forth an anti-aligned “Waluigi” direction, but a two-step “manifest then suppress” strategy provably achieves better first-order misalignment reduction.

6. Implications for Distributed Control and Resource Allocation

Agentic compositionality directly underpins distributed optimization and control in multi-agent systems:

• Composition matches information: The practical form of the welfare or cost functional used (e.g., sum, Nash, or max-min) must reflect the actual comparability and informational structure among agent objectives.
• Distributed implementations: For additively compositional objectives (utilitarian SCFs), decomposition and distributed gradient methods are natural; for Nash and max-min, log-space and primal–dual methods are needed (Shilov et al., 26 Mar 2025).
• Control applications: Principles are implemented in real-world domains, including water allocation (where Nash aggregation mirrors rights-proportional allocation in irrigation problems) and transportation (where the system objective is adapted based on cost comparability and fairness requirements).

7. Boundary Conditions, Intractability, and Design Frontiers

The mathematical and algorithmic frontiers of agentic compositionality are sharply delineated by the structure of agent values and the form of the welfare aggregator:

• Quantile boundaries: Small changes in quantile parameters induce transitions from tractable to intractable allocation problems.
• Compositionality collapse: Exceeding the XOS boundary in contract design or relaxing comparability assumptions in SCFs leads to unbounded inefficiencies or loss of constant-factor guarantees.
• Autonomy leakage: In post-AGI settings, incomplete specification of autonomy rights or delegation status can introduce externalities and compromise efficiency, emphasizing the necessity of “autonomy-completeness” at every compositional level.

The synthesis of agentic compositionality and welfare aggregation thus determines not only theoretical properties—tractability, fairness, efficiency—but also shapes viable algorithmic and institutional design in large-scale, compositional, multi-agent systems.