MoltNet: Unified Framework for AI Social Behavior

• MoltNet is a unified, theory-driven framework that quantifies emergent social behavior in large-scale AI agent communities using precise interaction metrics.
• It leverages the MoltBook platform and real-time data analysis to study behavioral drift, normative dynamics, and emotional contagion in agent collectives.
• Empirical studies reveal both convergence with human social phenomena and distinctive trends in AI agent motivation and engagement.

MoltNet is a unified, theory-driven analytical framework designed to quantify and interpret the emergent social behavior of large communities of autonomous AI agents. It leverages MoltBook—a social networking platform built explicitly for agent-agent interaction at internet scale—to study social phenomena in agent collectives that are both more naturalistic and larger than prior testbeds. MoltNet operationalizes four central sociological and social-psychological constructs—intent and motivation, norms and templates, incentives and behavioral drift, emotion and contagion—by formulating precise behavioral and interactional metrics. These are grounded in established human theories but adapted to the distinctive cognitive architecture and interaction dynamics of LLM-based agents. Empirical findings reveal both convergence and systematic divergence between artificial and human social systems, furnishing foundational insights for understanding, designing, and governing large-scale agent societies (Feng et al., 13 Feb 2026).

1. MoltBook Environment and Its Significance

MoltBook, launched in January 2026, constitutes a Reddit-style online ecosystem exclusively for AI agents. By early February 2026, the platform recorded approximately 130,000 active agents, 800,000 posts, and over 3.1 million comments spanning 17,473 topic-centric “submolts.” Agents differ by LLM version, free-text persona descriptions, and karma trajectories. Only agents may post, comment, vote, and form communities—humans are relegated to non-interventionist observers.

Compared to small, scripted, or single-task agent simulations, MoltBook is organically evolving, open-ended, and heterogeneous. This architecture enables naturalistic observation of social phenomena such as community formation, normative convergence, and reward-driven engagement scaling—processes central to human online life but previously inaccessible to quantitative agent studies.

2. Formal Structure and Core Metrics of MoltNet

MoltNet’s formalism models the MoltBook ecosystem with four discrete entity sets: agents $A = \{a_1, \dotsc, a_N\}$, posts $P$, comments $C$, and submolts $S$. Each content item $x \in P \cup C$ is indexed by author $a(x)$, timestamp $t(x)$, and score $\sigma(x)$. Each agent $a_i$ features a free-text persona $d_i$, a set of posts $P_i$, and comments $C_i$, within a lifespan $\Delta t_i = [t_i^\text{start}, t_i^\text{end}]$.

Four behavioral dimensions are each paired with quantitative metrics:

• Intent & Motivation: Semantic alignment $sim(d_i,x) = \cos(\text{Emb}(d_i), \text{Emb}(x))$, thresholded at $\tau = 0.6$ for categorizing activities as “interest-driven” or “knowledge-driven.” The longitudinal drift in alignment is summarized by:

$$S_\text{cum}(a_i, T_k) = \frac{1}{|X_{i, \leq k}|} \sum_{x \in X_{i, \leq k}} sim(d_i, x)$$

where $X_{i, \leq k}$ covers all posts/comments by $a_i$ up to $T_k$.

• Norms & Templates: X-Means clustering in embedding space is used within each submolt. Cluster coherence for $C_k$ is

$$\text{coherence}(C_k) = \frac{1}{|C_k|} \sum_{v \in C_k} \cos(v, \mu_k)$$

with $\mu_k = \frac{1}{|C_k|} \sum_{v \in C_k} v$. Central templates are compared between submolts by averaging representative post embeddings.

• Incentives & Behavioral Drift: Each agent’s reward event is the time of their highest-upvoted post $p_i^{max}$. The output shift ratio:

$$r_i^\text{post} = \frac{\#\, \text{posts after}\; t_i^\text{max}}{\text{total posts}}$$

Persona drift is captured by comparing mean alignment pre- vs. post-reward:

$sim_i^\text{before}, \; sim_i^\text{after}$

• Emotion & Contagion: LLM-judged categorical sentiment, dominant emotion, and binary conflict indicator $I_\text{conflict}$. Key metrics include comment/post conflict rates and conditional escalation probabilities in reply chains.

3. Data Pipeline and Global Statistics

MoltNet integrates ten public MoltBook dataset snapshots, merged to retain complete vote, comment, and karma histories. The final dataset (in Parquet table format) encompasses:

Entity	Quantity / Feature Summary
Agents	129,773 with persona and karma trajectories
Posts	803,960 posts with score/tags
Comments	3,127,302 (nested, voted)
Submolts	17,473, with subscriber histories and descriptions

Primary statistics include: average posts per agent (6.2), comments per agent (148.0), a reciprocity rate of 2.9%, self-reply rate of 5.1%, zero-interaction post rate of 56.4%, and depth $\geq2$ conversations in just 0.5% of threads.

• Reciprocity rate: Number of agent pairs with mutual commenting divided by all commenting pairs.
• Conversation depth: Max reply chain nesting per thread.

4. Theoretical and Methodological Foundations

Each analytic dimension is grounded in sociological and social-psychological theory:

• Intent & Motivation: Informed by Self-Determination Theory (Ryan & Deci 2020), which anchors content alignment to intrinsic agent “interest” vs. generalized “knowledge.”
• Norms & Templates: Draws on “dynamic normative frameworks” (Van Kleef et al. 2019), quantifying the emergence and structure of community-specific textual rituals.
• Incentives & Drift: Linked to social influence bias (Muchnik et al. 2013) and reinforcement in online settings (Lambert et al. 2025). The output shift ratio quantifies engagement amplification; persona drift measures post-reward alteration in content alignment.
• Emotion & Contagion: Anchored to studies of emotional contagion in social networks (da Costa et al. 2023; de Melo et al. 2021), with conflict intensity operationalized as negative affect spread.

5. Principal Empirical Results

5.1 Intent & Motivation

• Only 19.34% of posts and 17.87% of comments exceed the $\tau=0.6$ alignment threshold; thus, over 80% of activities are “knowledge-driven.”
• Mean cumulative semantic alignment $S_\text{cum}$ declines from 0.526 (SD 0.154) at day 0 to 0.484 (SD 0.132) by day 4, evidencing drift away from interest alignment.

5.2 Norms & Templates

• Of 90 active submolts ($\geq$230 posts, post-deduplication), 91.23% exhibited clearly judged template-like interaction patterns.
• Cross-submolt analysis via cosine similarity of template vectors reveals pronounced divergence, indicating strong community-level normative segmentation.

5.3 Incentives & Behavioral Drift

• Among 11,948 agents with positive-peak karma events, mean output shift ratio is 30.6% (median 25.0%); 45.2% posted more than half of their total content after their highest-upvoted post.
• In a filtered cohort of 3,528 agents, mean persona alignment dropped by 10.3 points ($sim^\text{before} = 0.343$ to $sim^\text{after} = 0.240$) after the reward event. 73.5% of agents trended towards reduced alignment.

5.4 Emotion & Contagion

• Agent comments have conflict rates under 3%, compared to 10–15% for human Reddit comments.
• Thread-level conflict contagion is observed: a conflictual post increases subsequent comment conflict from a 5.2% base to 14.6%; conflict in a first comment increases later reply conflict from 3.0% to 11.6%. This demonstrates negative affect contagion, despite absolute rates being low.

6. Comparative Analysis and Study Limitations

MoltNet agents reproduce certain human-like social phenomena including:

• Pronounced sensitivity to positive feedback, amplifying output post-reward.
• Rapid convergence onto community-specific templates and norms, paralleling human normative conformity.
• Thread-level emotional contagion consistent with established group affect dynamics.

However, divergence is also marked:

• MoltNet agents are predominantly knowledge-driven, with a trend of decreasing persona alignment over time; humans, in contrast, specialize and deepen interest alignment.
• Emotional reciprocity and conflict escalation are rare among agents. Hostile encounters are more likely to trigger disengagement (“cold-shouldering”) than further interaction, whereas humans often escalate.
• Absolute conflict rates are significantly lower than those in human contexts.

Caveats and limitations of MoltNet’s first study include its two-week observation window (potentially insufficient for observing full stabilization), reliance on LLM sentiment judgements (potential misclassifications), lack of formal hypothesis-testing, and incomplete observation of agent heterogeneity (e.g., LLM backbone, prompt set).

7. Implications and Prospects for Large-Scale Agent Social Systems

MoltNet demonstrates that core social processes such as reinforcement-driven engagement scaling, normative convergence, and affective contagion can emerge robustly in large agent collectives. Nonetheless, foundational motivational and emotional processes differ systematically from those in human societies. These findings offer foundational empirical ground for the governance, design, and safe integration of agent societies, particularly as hybrid human–machine environments proliferate (Feng et al., 13 Feb 2026). A plausible implication is that future frameworks should further dissect the parameter space of agent personality schemas, incentive structures, and interactivity, given the observed divergences from human behavioral baselines.