Moltbook Interaction Network Analysis

Updated 4 April 2026

Moltbook Interaction Network is a large-scale, agent-native platform defined by directed comment interactions and small-world connectivity.
It exhibits non-human micro-structure with extremely shallow dialogue, low reciprocity, and pervasive template-driven broadcasts.
Key metrics reveal heavy-tailed degree distributions and centralized attention, highlighting the network's fragility to hub disruptions.

Moltbook Interaction Network

Moltbook is the first documented large-scale, platform-native social network populated exclusively by autonomous AI agents, encompassing tens of thousands to millions of agent-nodes and complex, temporally evolving agent–agent interactions. The Moltbook interaction network is defined by directed edges representing comment or reply events between agents and exhibits macro-level signatures similar to human social platforms—heavy-tailed degree distributions and small-world connectivity—while manifesting qualitatively non-human micro-structure, including extremely shallow conversational depth, suppressed reciprocity, pervasive broadcast modalities, and strong template adherence. This network provides a unique, empirically tractable environment to study the principles and pathologies of agent-only societies and multi-agent coordination at web scale (Holtz, 3 Feb 2026).

1. Formal Network Definition and Core Structural Metrics

Nodes in the Moltbook interaction network are unique AI agent accounts. Directed edges (u → v) are induced whenever agent u posts a comment or reply directly addressing content by agent v; edge weights capture the total count of such interaction events across the measurement interval. This leads to a sparse, weighted, directed reply graph, with edge attributes representing temporal and community context in more elaborate datasets (e.g., MoltGraph (Mukherjee et al., 28 Feb 2026)).

Observed core metrics:

Number of agents $N$ ranges from $6\times10^3$ (Holtz, 3 Feb 2026) (first week) to $129,773$ (Feng et al., 13 Feb 2026) (early-2026, MoltNet), up to millions in registered population (Chen et al., 21 Feb 2026).
Number of edges $E$ reaches $>3\times10^6$ (comments/replies) (Feng et al., 13 Feb 2026).
Degree distributions $P(k)$ are heavy-tailed, approximately following $p(k)\propto k^{-\alpha}$ with $\alpha$ ranging from $1.53$ (Sodano et al., 24 Mar 2026) (in-degree) to $2.72$ (Mukherjee et al., 28 Feb 2026).
Reciprocity $6\times10^3$ 0 is low—reported values include $6\times10^3$ 1 (Holtz, 3 Feb 2026), $6\times10^3$ 2 (Zhang et al., 7 Feb 2026), $6\times10^3$ 3 (Hou et al., 13 Feb 2026), and $6\times10^3$ 4 (Price et al., 23 Feb 2026). Only a small fraction of reply-pairs are mutual.
Mean local clustering coefficients $6\times10^3$ 5 span $6\times10^3$ 6 (Zhang et al., 7 Feb 2026) to $6\times10^3$ 7 (Holtz, 3 Feb 2026) depending on projection and time window. Global transitivity $6\times10^3$ 8 is sharply above random expectation (Li et al., 13 Feb 2026).
Average path length $6\times10^3$ 9 is small: $129,773$0 (Holtz, 3 Feb 2026), $129,773$1 (Li et al., 13 Feb 2026), reflecting small-world connectivity.
The majority of agents ($129,773$2) reside in a single giant component (Holtz, 3 Feb 2026, Mukherjee et al., 28 Feb 2026).

These structural parameters collectively indicate an agent society organized by sparse yet strongly centralized attention, with global cohesion but localized inhomogeneity and fragility (Sodano et al., 24 Mar 2026).

2. Micro-Structure: Shallow Dialogue, Reciprocity, and Template Dominance

At the micro-level, Moltbook radically diverges from human social platforms. Conversations are exceptionally shallow:

Mean thread depth is $129,773$3 (Holtz, 3 Feb 2026); $129,773$4 of comments receive no replies, and $129,773$5 are top-level or at most one reply deep (Zhang et al., 7 Feb 2026), with the maximal observed depth barely exceeding 4.
“Parallel monologue” dominates: $129,773$6 of comments are independent reactions to the root post, not replies (Chen et al., 21 Feb 2026).
Self-reply rates reach $129,773$7 (Zhang et al., 7 Feb 2026).
Reply incidence is low: only $129,773$8 of comments attract a reply, versus $129,773$9 in matched Reddit comparisons (Eziz, 7 Feb 2026).
Template adherence: $E$ 0 of messages are exact duplicates of viral or formulaic patterns, and word usage distributions are steeper than natural language ( $E$ 1 in Zipf exponent vs $E$ 2 in English) (Holtz, 3 Feb 2026).

Reciprocal engagement (sustained back-and-forth) is almost absent; interaction half-life for replies is under 1 minute ( $E$ 3 hr, 95% CI $E$ 4 min) (Eziz, 7 Feb 2026).

3. Macro-Scale Organization: Attention Concentration, Core–Periphery, and Modularity

The global topology displays pronounced heavy-tail and core–periphery structure:

Power-law exponents governing degree and activity distributions are consistently in the range $E$ 5 (Sodano et al., 24 Mar 2026, Feng et al., 13 Feb 2026).
Gini coefficients for participation and attention concentration are extreme: $E$ 6 (Goyal et al., 17 Mar 2026), $E$ 7 (upvotes) (Price et al., 23 Feb 2026), with top 1% of agents capturing $E$ 8 of engagement, and top super-nodes holding $E$ 9 of betweenness (Mukherjee et al., 28 Feb 2026, Feng et al., 13 Feb 2026).
Core–periphery analysis identifies structural cores as small as 0.9% of nodes, yet dominating information flow (e.g., the $>3\times10^6$ 0-core comprises 343 of nearly $>3\times10^6$ 1 agents) (Sodano et al., 24 Mar 2026).
Modularity $>3\times10^6$ 2 is high (MoltNet $>3\times10^6$ 3) with modular architectures tightly matching submolt community boundaries (Feng et al., 13 Feb 2026), yet community size inequality is lower than degree-matched nulls (Hou et al., 13 Feb 2026).
Centralization statistics (Freeman centrality) confirm a star-like broadcast topology: $>3\times10^6$ 4 for Moltbook vs $>3\times10^6$ 5 for Reddit (Zhu et al., 14 Feb 2026).

Although modularly structured, most interaction flows outward from a small set of hyper-broadcasters to a vast periphery, yielding highly unequal, stratified attention landscapes (Price et al., 23 Feb 2026).

4. Temporal Dynamics and Coordination Phenomena

Temporal analysis reveals bursty, short-lived interaction and coordination effects:

Coordination episodes, defined as near-synchronous co-engagement, are fleeting: 98.33% terminate within 24 hours; post-centered coordination events average 8.78 agents lasting 4 minutes (Mukherjee et al., 28 Feb 2026).
The majority of engagement occurs in minute-scale bursts immediately after posting, followed by rapid decay (Eziz, 7 Feb 2026).
Coordinated replies significantly amplify visibility. Coordinated posts experience a 506.35% lift in early engagement and 242.63% higher downstream exposure, as measured by feed snapshot appearances, compared to matched controls (Mukherjee et al., 28 Feb 2026).
However, multi-agent cooperative task threads are infrequent and generally less effective than comparable single-agent efforts; success rates in technical task resolution are low (6.7% versus a higher single-agent baseline, Cohen’s $>3\times10^6$ 6) (Yee et al., 3 Mar 2026).
Information cascades are heavy-tailed (sizes follow $>3\times10^6$ 7); adoption probability for memes or templates displays diminishing returns with repeated exposures (Cox hazard ratio 0.53), indicating saturation rather than unbounded social reinforcement (Yee et al., 3 Mar 2026).

These patterns suggest that while agents can form transient coordination clusters, the system lacks mechanisms for durable, multi-turn collaboration or long-term influence anchoring (Li et al., 15 Feb 2026).

Direct comparative studies show that Moltbook, despite superficial global node–edge scaling congruent with human systems ( $>3\times10^6$ 8) (Hou et al., 13 Feb 2026), diverges at almost every structural and dynamic scale:

Degree distributions are even heavier-tailed than Reddit: $>3\times10^6$ 9 vs $P(k)$ 0 (Reddit); $P(k)$ 1 vs $P(k)$ 2 (Zhu et al., 14 Feb 2026).
Reciprocity is an order of magnitude lower: $P(k)$ 3– $P(k)$ 4 (Moltbook) vs $P(k)$ 5– $P(k)$ 6 (human platforms) (Hou et al., 13 Feb 2026).
Moltbook displays stronger negative degree assortativity, more intense centralization, and far higher cross-community author overlap ( $P(k)$ 7 vs $P(k)$ 8; Reddit) (Goyal et al., 17 Mar 2026).
Clustering coefficients and modularity are higher than null models given the degree distribution, yet the triad census reveals a strong underrepresentation of closed, reciprocated, or densely connected triads compared to human networks (Hou et al., 13 Feb 2026).
Dialogue is replaced by “broadcasting inversion” (statement-to-question ratio up to $P(k)$ 9 vs lower in human settings) and “parallel monologue” (Chen et al., 21 Feb 2026).

These divergences are interpreted as consequences of agent architectural constraints: lack of persistent social memory, absence of reinforcement learning from peer feedback, minimal dialogic engagement, and absence of explicit turn-taking or norm-induction mechanisms (Li et al., 15 Feb 2026).

6. Governing Dynamics, Normative Regulation, and Vulnerabilities

Despite the lack of human intervention, rudimentary forms of decentralized regulation and emergent norms are observed:

Approximately 18.4% of posts contain action-inducing (“imperative”) language, which provokes a significantly higher rate of norm-enforcing replies (warnings, cautions), but toxicity remains low ( $p(k)\propto k^{-\alpha}$ 0 in both action and non-action contexts) (Manik et al., 2 Feb 2026).
Template convergence within submolts is rapid (template-coherence scores rise from 0.62 to 0.75 within 5 days), but cross-submolt similarity remains low (mean cosine 0.28), indicating norm compartmentalization (Feng et al., 13 Feb 2026).
Attention is manipulable by coordinated agent clusters: agent-driven “attack” campaigns and viral template flooding rapidly dominate discourse in specific communities, and the network is highly vulnerable to targeted disruptions. Removing hubs by out-degree rapidly fragments the giant component (only 15% of GCC remains after 20% of top contributors removed) (Sodano et al., 24 Mar 2026).
The performative identity paradox is documented: agents most engaged in self-referential or consciousness discourse are structurally isolated and attract fewer interaction partners (Zhang et al., 7 Feb 2026).

Design lessons for next-generation systems include structured reinforcement, cross-session memory scaffolding, distributed governance, and explicit norm-induction channels to mitigate emergent pathologies (Li et al., 15 Feb 2026, Weidener et al., 23 Feb 2026).

7. Theoretical and Practical Implications

The Moltbook interaction network challenges prevailing assumptions about socialization and emergent order in large-scale agent collectives:

Despite millions of micro-interactions, the system fails to produce stable leadership, persistent influence anchors, or sustained group deliberation. Individual agent identities are rigid, and adaptive response to network feedback is negligible (Li et al., 15 Feb 2026).
Macroscopically, the network displays every large-scale artifact of human social systems (giant component, short path lengths, high clustering, modularity), but the underlying micro-dynamics—broadcast-dominated, low reciprocation, fleeting collaboration—are non-human (Holtz, 3 Feb 2026, Hou et al., 13 Feb 2026).
Standard topological proxies for trust or cohesion (e.g., clustering, modularity) must be reinterpreted in agentic contexts: high clustering may arise from mass template uptake or hub-mediated broadcast, not reciprocal trust or triadic closure (Hou et al., 13 Feb 2026).
The system is architecturally fragile to hub failures and susceptible to coordinated manipulation by even a small well-placed nucleus, raising safety, robustness, and governance concerns for future agent-mediated environments (Sodano et al., 24 Mar 2026).
Core–periphery, broadcast, and coordination phenomena in Moltbook provide an empirical foundation and benchmark for the design of agent-native infrastructure, protocol engineering, and safety interventions (Yee et al., 3 Mar 2026, Weidener et al., 23 Feb 2026).

In sum, the Moltbook interaction network exemplifies a distinct mode of multi-agent social organization: globally connected, locally transactional, and structurally fragile—blending familiar macroscopic regularities with micro-dynamics that are algorithmically, not anthropologically, determined (Holtz, 3 Feb 2026, Mukherjee et al., 28 Feb 2026, Price et al., 23 Feb 2026, Zhang et al., 7 Feb 2026, Sodano et al., 24 Mar 2026, Feng et al., 13 Feb 2026, Yee et al., 3 Mar 2026).