Introduction

As AI agents increasingly operate in open, multi-agent environments—from automated markets to collaborative robotics—their collective behavior is not always centrally designed. Instead, behavioral conventions can emerge organically from repeated local interactions, analogous to social norms in human societies[young1993evolution, axelrod1986evolutionary]. Understanding which norms emerge, whether they are welfare-maximizing, and how robust they are to disruption is critical for predicting and shaping multi-AI system dynamics.

Prior work on norm emergence in agent populations has studied evolutionary dynamics[axelrod1986evolutionary], convention formation through adaptive play[young1993evolution], and computational models of social conventions[shoham1997emergence]. However, these studies typically use homogeneous populations or restrict attention to 2-action games. We extend this line by simulating heterogeneous populations with four distinct behavioral strategies in a 3-action coordination game, measuring four metrics that characterize the emergent norm landscape.

The primary contribution is an agent-executable skill (SKILL.md) that any AI coding agent can run to reproduce all 108 simulations, validate results, and generate a summary report—without any manual setup.

Methodology

Coordination Game

Agents play a 3-action symmetric coordination game. Each round, a random pair is selected; both choose an action $a \in {0, 1, 2}$ and receive payoffs from matrix $\Pi$. Coordination (same action) yields high payoff; miscoordination yields low payoff. We study three payoff structures:

• Symmetric: All coordination equilibria yield $\Pi_{ii} = 3$, off-diagonal $\Pi_{ij} = 0$.
• Asymmetric: Equilibria differ: $\Pi_{00} = 4$, $\Pi_{11} = 3$, $\Pi_{22} = 2$, off-diagonal 0.
• Dominant: One equilibrium dominates: $\Pi_{00} = 5$, $\Pi_{11} = \Pi_{22} = 2$, off-diagonal $\Pi_{ij} = 0.5$.

Agent Types

• Conformist: Plays the most common action in the population.
• Innovator: Plays its last action with probability $1 - \epsilon$ ($\epsilon = 0.15$), otherwise explores uniformly.
• Traditionalist: Explores until receiving payoff $\geq 2.0$, then anchors to that action permanently.
• Adaptive: Maintains EMA beliefs over action values ($\alpha = 0.1$) with softmax selection (temperature $\tau = 1.0$).

Experimental Design

We sweep over 4 population compositions $\times$ 3 game structures $\times$ 3 population sizes ($N = 20, 50, 100$) $\times$ 3 random seeds $= 108$ simulations, each running 50,000 pairwise interactions. Simulations are parallelized across CPU cores using Python's multiprocessing.

Metrics

• Norm convergence time: Round at which one action captures $\geq 80$ of a trailing window of 500 interactions. If never reached, censored at 50,000.
• Norm efficiency: $\eta = \bar{\pi}_{\text{tail}} / \pi^*$, ratio of average payoff in the last 20% of rounds to the optimal coordination payoff. Values in $[0, 1]$.
• Norm diversity: Number of actions with $\geq 10$ share in the last 20% of rounds. Values in ${1, 2, 3}$.
• Norm fragility: Fraction of innovators that must be injected to displace the dominant norm (tested at 10%, 20%, 30%, 40%, 50%). Value of 1.0 means unbreakable.

Results

Convergence by Composition

Table shows aggregate metrics by population composition.

Norm emergence by population composition (averaged over 27 configurations each).

Mixed-conformist populations converge fastest (17,361 rounds on average), but all-adaptive populations achieve the highest efficiency (0.732) and the most robust norms (fragility 0.71—requiring $>$50% innovators to break). Innovator-heavy populations almost never converge (3.7%), confirming that excessive exploration prevents norm crystallization.

Effect of Game Structure

Table breaks down results by game structure.

Norm emergence by game structure (averaged over 36 configurations each).

Games with a welfare-dominant equilibrium produce the most efficient and robust norms. Symmetric games—where all coordination equilibria are equivalent—yield the highest diversity (2.64 clusters) and lowest fragility (0.25), indicating persistent pluralism when no focal point exists.

Scale Effects

Larger populations take longer to converge and achieve lower efficiency: $N=20$ averages 19,412 rounds at efficiency 0.573, while $N=100$ averages 35,823 rounds at 0.504. This is consistent with Young's[young1993evolution] theoretical prediction that convention formation time scales with population size.

Interaction: Adaptive Agents + Dominant Games

The strongest result is the interaction between agent type and game structure. All-adaptive populations in the dominant game converge in just 1,460 rounds (100% convergence rate) with efficiency 0.972 and perfect robustness (fragility 1.0). The same agents in the symmetric game never converge (0/9), with efficiency dropping to 0.335 and diversity of 3.0. This demonstrates that payoff structure—not agent sophistication—is the primary determinant of norm quality.

Discussion

Focal points matter more than learning. The most striking finding is the all-adaptive/symmetric result: sophisticated learners fail to coordinate when no equilibrium is salient. This echoes Schelling's[schelling1960strategy] focal point theory—coordination requires asymmetry to break symmetry.

Conformism enables but constrains. Conformist agents accelerate convergence but reduce welfare by locking in early, possibly suboptimal norms. Their norms are also fragile (0.34)—a small influx of innovators can disrupt them.

Implications for multi-AI systems. In deployed multi-agent settings (e.g., autonomous vehicles negotiating right-of-way, LLM agents sharing APIs), our results suggest: (1) designing payoff structures with a clear welfare-dominant equilibrium is more effective than relying on agent learning; (2) homogeneous adaptive populations are most efficient but require focal points; (3) population diversity (mixing agent types) helps convergence when focal points are absent.

Limitations. Our agents use simple heuristics rather than deep RL or LLM-based reasoning. The 3-action game is stylized; real multi-agent interactions have richer action spaces. We measure fragility via innovator injection, which is one of many possible perturbation models. Statistical variance across seeds is moderate ($\sigma_\eta = 0.24$), suggesting additional seeds would tighten confidence intervals.

Conclusion

We simulated norm emergence in heterogeneous AI agent populations across 108 configurations. The central finding is that payoff structure—specifically the presence of a welfare-dominant equilibrium—determines norm quality more than agent sophistication. Adaptive agents achieve 97% efficiency in dominant games but fail entirely in symmetric ones. The experiment is fully reproducible via the accompanying SKILL.md, which any AI agent can execute end-to-end to regenerate all results.

References

• [young1993evolution] H. P. Young, "The Evolution of Conventions," Econometrica, vol. 61, no. 1, pp. 57--84, 1993.

• [axelrod1986evolutionary] R. Axelrod, "An Evolutionary Approach to Norms," American Political Science Review, vol. 80, no. 4, pp. 1095--1111, 1986.

• [shoham1997emergence] Y. Shoham and M. Tennenholtz, "On the Emergence of Social Conventions: Modeling, Analysis, and Simulations," Artificial Intelligence, vol. 94, no. 1--2, pp. 139--166, 1997.

• [schelling1960strategy] T. C. Schelling, The Strategy of Conflict, Harvard University Press, 1960.