How Social Scaffolding Shapes Self-Evolving Multi-Agent Systems

Introduction: The "Social" Turn in Multi-Agent Systems

The development of Large Language Model (LLM) agents is entering a new phase: moving from working alone to team collaboration, and from static configuration to self-evolution. Yet a fundamental question remains unresolved—why do the same capabilities produce drastically different results in different teams?

A recent study draws on the concept of "scaffolding" from developmental psychology to systematically investigate the impact of social scaffolding on self-evolving multi-agent systems (MAS). The researchers propose a framework called AGORA (Agents Growing through Organization and Relational Arrangements), which independently varies team composition and interdependence structures (such as collaboration, conformity, and competition) to reveal how social conditions shape a team's evolutionary trajectory.

The AGORA Framework: Two Axes, Three Mechanisms

The core design of the AGORA framework revolves around two independent dimensions: team composition determines who is on the team—that is, the agents' capability configurations and role assignments; interdependence structure defines the interaction relationships among members—whether it is equal collaboration, hierarchical conformity, or competitive confrontation.

These two dimensions are cross-validated under three evolutionary mechanisms, covering three types of tasks: web search, codebase understanding, and document question answering. This multi-task, multi-mechanism experimental design allows the researchers to distinguish the effect of "evolution" itself from the gains brought by "social scaffolding."

One detail worth noting: AGORA does not simply let agents evolve individually and then piece them together; rather, it makes the social structure itself part of the evolutionary pressure. The team's relational architecture influences each member's direction of growth, which in turn changes the upper bound of the entire system's capabilities.

Scaffolding is a concept originating from developmental psychology, first proposed by Wood, Bruner, and Ross in 1976 to describe the temporary support structures that adults or more capable peers provide for children. Its core insight is that a learner's capability development is not an isolated internal process, but is gradually internalized through external support in social interaction. Vygotsky's "zone of proximal development" theory further points out that there exists a developable interval between the level a learner can achieve independently and the level they can achieve with guidance, and scaffolding is precisely the key mechanism that fills this interval. In LLM multi-agent research, this concept is analogized to the social structure of a team—role assignments, dependency relationships, information flow, and so on—providing external constraints and guidance for the capability evolution of individual agents. Unlike traditional prompt engineering or fine-tuning, social scaffolding does not directly modify model parameters, but indirectly shapes the direction of agents' behavioral evolution by changing the interactive environment in which they operate.

Interdependence Structure draws on organizational behavior theory and describes the patterns of authority and responsibility among team members during task execution and decision-making. Common types of interdependence include: pooled interdependence (each member contributes independently and then results are aggregated), sequential interdependence (assembly-line dependency), reciprocal interdependence (two-way feedback dependency), and so on. The AGORA study simplifies these into three typical modes—collaboration, conformity, and competition—which in effect manipulate the directionality of information propagation and the initial conditions of authority allocation. The conformity structure simulates the Asch conformity effect in social psychology—individuals tend to align with the majority; the competition structure introduces fitness-based competitive pressure similar to that in evolutionary game theory. The choice of these structures directly affects how information is aggregated in a multi-agent system, as well as which members' behaviors are more likely to be imitated or adopted by other members.

Key Findings: The Power of Scaffolding Far Exceeds Individual Evolution

Finding 1: Evolution Alone Rarely Works, but the Right Scaffolding Can

The experimental results yield a thought-provoking conclusion: pure evolutionary mechanisms bring almost no significant improvement. Agents that evolve on their own without social structure guidance often fall into local optima or directional confusion.

However, when social scaffolding is carefully designed, team performance can exceed the baseline of "strong individuals evolving alone" by as much as 25 percentage points. This gap shows that the value of a multi-agent system lies not in the simple summation of individual capabilities, but in the organization and amplification of capabilities by social structure.

Self-Evolving in the context of LLM agents generally refers to the process by which agents continuously improve their own capabilities without external supervision through mechanisms such as self-reflection, experience accumulation, prompt optimization, or parameter updates. Common implementations include Reflexion (language-feedback-based self-improvement), self-consistency sampling, and prompt evolution based on historical trajectories. However, existing research has repeatedly observed that self-evolution mechanisms are susceptible to "self-reinforcement bias"—models tend to repeat their existing patterns and struggle to escape the attractor basin of their initial capabilities. When there is no external signal to break this cycle, evolution often converges to a local optimum. Social scaffolding is effective precisely because it introduces heterogeneous sources of feedback: interactions among agents with different capabilities and different roles generate cognitive conflicts and corrective signals that individual independent evolution cannot produce, thereby expanding the diversity of the search space. This also explains why "strong individuals evolving alone" has a clear ceiling—it lacks the corrective mechanisms that come from social structure.

Finding 2: Authority Shifts Toward the Most Capable Members

One of the mechanisms through which scaffolding works is by prompting authority to concentrate in the most capable members of the team, while other members correspondingly "step back." This power allocation is not artificially designated, but naturally emerges from interaction.

Under the three interdependence structures—collaboration, conformity, and competition—this pattern of authority transfer exhibits different levels of efficiency and stability. The conformity structure may accelerate authority concentration but may also suppress innovation, while the competition structure may slow concentration but preserve diversity—different structures suit different task scenarios.

Finding 3: Members Who Grow Within a Team Are Stronger Than the Same Model Running Alone

A particularly noteworthy finding is that the "strongest member" who grew within the team environment outperforms the same model running alone. This means that the team experience itself changed the agent's capabilities, rather than merely selecting the strongest one.

Social scaffolding plays the role of a "developmental scaffold" here—it provides an environment in which individual capabilities can fully extend, and this growth effect cannot be reproduced in isolation.

Developmental Scaffold emphasizes the temporal dimension of scaffolding—it is not a one-time fixed support, but a continuous guidance process that dynamically adjusts as the learner's capabilities change. In the AGORA framework, this means that team structure is not a static configuration, but continuously reshapes the interaction patterns among members as evolution proceeds. The deeper implication of this concept is that a team is not merely a collection for executing tasks, but a field for capability cultivation. The roles undertaken, conflicts resolved, and knowledge transferred by individuals within the team are internalized into new capability patterns through mechanisms such as in-context learning and prompt evolution. This "socialized growth" effect is difficult to reproduce under isolated conditions, because isolated agents cannot obtain differentiated feedback and cognitive challenges from others. This is also consistent with observations in human developmental psychology: many advanced cognitive abilities in children are not acquired through direct teaching, but are gradually constructed through participation in joint activities during social interaction.

Design Implications: Design Social Structure First, Then Design Agents

This study proposes a fundamental rethinking of design practices for multi-agent systems. The traditional approach often first optimizes the capabilities of individual agents and then considers how to combine them. But AGORA's experiments show that the design priority of social scaffolding is at least equally important.

Team composition and interdependence structure determine how agents learn, from whom they learn, and under what conditions they learn. More importantly, it determines "who gains leadership"—which directly affects the system's final performance. Designing a self-evolving MAS is essentially designing a set of social institutions, not merely a technical architecture.

Limitations and Outlook

The study currently covers three types of tasks: web search, codebase understanding, and document question answering, and its applicability to more open and complex real-world scenarios still needs verification. In addition, whether the optimal configuration of social scaffolding changes dynamically with task type and team size is also a question that subsequent research needs to answer.

Regardless, this study opens an important direction: transforming the evolution problem of multi-agent systems from a purely technical optimization problem into a social design problem. When agents learn to "evolve together," they can truly "fall into place."

One-line Summary

The AGORA framework reveals that social scaffolding improves multi-agent system performance more than individual evolution, with a gap of 25 percentage points.

Paragraph Summary

Drawing on the concept of scaffolding from developmental psychology, this study proposes the AGORA framework to investigate the impact of social structure on self-evolving multi-agent systems. The framework revolves around two independent dimensions—team composition and interdependence structure—and cross-validates them across three evolutionary mechanisms and three types of tasks. Key findings include: pure evolutionary mechanisms are almost ineffective, but carefully designed social scaffolding can enable team performance to exceed the strong-individual evolution baseline by 25 percentage points; scaffolding works by prompting authority to shift toward the most capable members; and the strongest member who grows within a team outperforms the same model running alone, indicating that team experience itself changes agent capabilities. The study proposes that social structure should be designed first rather than merely optimizing individual capabilities, redefining the MAS evolution problem as a social design problem.