Abstract
As artificial intelligence (AI) enters the agentic era, large language models (LLMs) are increasingly deployed as autonomous agents that interact with one another rather than operate in isolation. This shift raises a fundamental question: how do machine agents behave in interdependent environments where outcomes depend not only on their own choices but also on the coordinated expectations of peers? To address this question, we study LLM agents in a canonical network-effect game, where economic theory predicts convergence to a fulfilled expectation equilibrium (FEE). We design an experimental framework in which 50 heterogeneous GPT-5-based agents repeatedly interact under systematically varied network-effect strengths, price trajectories, and decision-history lengths. The results reveal that LLM agents systematically diverge from FEE: they underestimate participation at low prices, overestimate at high prices, and sustain persistent dispersion. Crucially, the way history is structured emerges as a design lever. Simple monotonic histories—where past outcomes follow a steady upward or downward trend—help stabilize coordination, whereas nonmonotonic histories amplify divergence and path dependence. Regression analyses at the individual level further show that price is the dominant driver of deviation, history moderates this effect, and network effects amplify contextual distortions. Together, these findings advance machine behavior research by providing the first systematic evidence on multi-agent AI systems under network effects and offer guidance for configuring such systems in practice.
Introduction
As AI transitions from isolated chatbots to "agentic" systems capable of autonomous action—such as trading stocks, managing logistics, or participating in digital marketplaces—a new frontier of machine behavior emerges. These agents do not operate in a vacuum; they inhabit environments governed by network effects, where the value of an action depends on how many other agents choose the same path.
In classical economics, such scenarios are analyzed through the lens of the Fulfilled Expectation Equilibrium (FEE). This theory assumes that rational actors will eventually converge on a shared, accurate expectation of peer participation. However, Large Language Models (LLMs) do not "solve" for equilibrium through mathematical optimization. Instead, they predict the next token based on learned patterns and provided context. This paper investigates whether a community of LLM agents can achieve economic coordination or if the inherent "machine logic" of transformers leads to systematic market failure.
Experimental Framework
To test these dynamics, we constructed a canonical network-effect game involving 50 heterogeneous agents. Each agent was tasked with deciding whether to "attend a conference" based on three factors: a private valuation, the current price, and a "network effect" bonus determined by the total number of attendees.
Our study utilized a $2 \times (1 + 3 \times 4)$ experimental design to isolate key variables:
- Model Architectures: The primary simulations used GPT-5, with Qwen3-Plus serving as a robustness check.
- Network Strength ($\beta$): We tested both weak (0.25) and strong (0.75) interdependencies.
- Temporal Scenarios: We compared Static (single-shot) decisions against Dynamic (repeated) rounds where agents could observe past outcomes.
- History Length: Agents were provided with 1, 3, or 6 rounds of previous participation data.
- Price Trajectories: We manipulated the price over time across four paths—increasing, decreasing, converging (moving toward the mean), and diverging (moving toward extremes)—to observe how agents learn from trends.
The entire simulation workflow was formalized as a Finite State Machine (FSM), ensuring that information flow and timing were strictly controlled and reproducible. This generated a dataset of 7,800 individual observations.
Key Insights: The Machine Logic of History
1. History is Not "Sunk"
In traditional economic models, the past is often considered a "sunk cost"—irrelevant to a rational, forward-looking actor. For LLMs, however, history is the primary "raw material" for reasoning. We found that an agent's beliefs are fundamentally anchored to the sequence of tokens they have previously processed. The past is not just context; it is the foundation of their predictive capability.
2. Systematic Behavioral Biases
LLM agents do not behave as perfectly rational economic actors. They exhibit a consistent pattern of distortion: they systematically underestimate participation when prices are low and overestimate it when prices are high. This suggests that LLMs may be overly sensitive to price signals, treating them as proxies for popularity or quality in ways that deviate from the mathematical equilibrium.
3. History as a Design Lever
The structure of history proved more influential than the volume of history.
- Monotonic Histories: When past outcomes followed a clear upward or downward trend, they acted as stabilizing cues. These trends helped agents align their expectations and significantly reduced coordination errors.
- Non-Monotonic Histories: Oscillating or complex patterns (such as converging or diverging price paths) "scrambled" the predictive structure of the LLM. This led to persistent dispersion in expectations and a breakdown in coordination, which we term the "complexity penalty."
4. The Amplification Effect of Networks
Network effects do not act as an independent source of error. Instead, they function as a catalyst. Stronger network interdependencies (high $\beta$) magnified the existing biases caused by price fluctuations and agent heterogeneity, making the collective "machine rationality" much more fragile.
Quantitative Results
Our analysis quantified the gap between LLM behavior and theoretical FEE using Root Mean Squared Error (RMSE):
- Equilibrium Divergence: In static settings, the RMSE was 6.712 under weak network effects but nearly doubled to 13.979 under strong network effects.
- The Power of Trends: In monotonic price sequences, increasing the history window from 1 to 13 rounds dramatically improved coordination, slashing the RMSE from approximately 9.0 to 2.5.
- The Complexity Penalty: In non-monotonic scenarios, even a long history window failed to bring agents close to equilibrium, with RMSE remaining stubbornly high (often above 9.0).
- Regression Analysis: Individual-level regressions identified price as the dominant driver of deviation. The interaction between Network Effects and Price was highly significant ($\beta = 14.294$), confirming that strong interdependencies exacerbate contextual distortions.
Conclusion and Future Work
This research provides a cautionary tale for the designers of multi-agent systems. The ability of an AI collective to reach a stable, rational equilibrium is not a given; it is a property that must be engineered. By carefully scaffolding the information and history provided to agents, designers can steer these systems toward better coordination.
Future research should address the limitations of this study, including the stylized nature of the "conference game" and the specificity of Transformer architectures. Testing different model families, such as State Space Models, and exploring adaptive prompting techniques will be essential as we move toward a world where machines increasingly manage our markets and social infrastructures.
Summary updated