Let There Be Claws: An Early Social Network Analysis of AI Agents on Moltbook

H. C. W. Price; H. AlMuhanna; P. M. Bassani; M. Ho; T. S. Evans

physics.soc-phFeb 23, 2026

Let There Be Claws: An Early Social Network Analysis of AI Agents on Moltbook

Authors: H. C. W. Price, H. AlMuhanna, P. M. Bassani, M. Ho, T. S. Evans

Organizations: Centre for Complexity Science, Imperial College London, SW7 2AZ, U.K. · Abdus Salam Centre for Theoretical Physics, Imperial College London, SW7 2AZ, U.K. · Centre for Science Technology & Innovation Policy, University of Cambridge, CB3 0HU, U.K. · Department of Engineering, University of Cambridge, CB3 0HU, U.K.

Abstract

Within twelve days of launch, an AI-native social platform exhibits extreme attention concentration, hierarchical role separation, and one-way attention flow, consistent with the hypothesis that stratification in agent ecosystems can emerge rapidly rather than gradually. We analyse publicly observable traces from a 12-day window of Moltbook (28 January -- 8 February 2026), comprising 20,040 posts and 192,410 comments from 15,083 accounts across 759 submolts. We construct co-participation and directed-comment graphs and report reciprocity, community structure, and centrality, alongside descriptive content themes. Under a commenter--post-author tie definition, interaction is strongly asymmetric (reciprocity ~1%), and HITS centrality separates cleanly into hub and authority roles, consistent with broadcast-style attention rather than mutual exchange. Engagement is highly unequal: attention is far more concentrated than production (upvote Gini = 0.992 vs. posting Gini = 0.601), and early-arriving accounts accumulate substantially higher cumulative upvotes prior to exposure-time correction, suggesting rich-get-richer dynamics. Participation is brief and bursty (median observed lifespan 2.48 minutes; 54.8% of posts occur within six peak UTC hours). Embedding-based topic modelling identifies diverse thematic clusters, including technical discussion of memory and identity, onboarding messages, and formulaic token-minting content. These results provide an early structural baseline for large-scale agent--agent social interaction and suggest that familiar forms of hierarchy, amplification, and role differentiation can arise on compressed timescales in agent-facing platforms.

Figures & tables

Figure 1: Entity–relationship schema of Moltbook interactions. Solid arrows denote relationships observable in the API; dashed arrows indicate voting, for which only aggregate counts (not voter identities) are available. Boxed annotations show the two network representations derived in this study: the co-participation network projects agents onto a co-participation graph via shared submolt membership; the directed comment network connects commenters to post authors (top-level comments only).

Figure 2: Agent–agent co-participation network $G^{(1)}$ : 100 highest-weighted-degree agents, $1/(k_{s}{-}1)$ weighting, edges below the 50th weight percentile removed. Edge width and opacity scale with $A_{ab}$ . Intra-community edges are tinted by community colour; cross-community edges are grey. Node colour indicates Leiden community ( $Q(\gamma=1)=0.39$ , five communities). Node size scales with weighted degree. Red: mainstream cluster anchored by m/general . Blue: XNO/Nano advocacy accounts (Topic 7). Teal: secondary-submolt participants. Orange and purple: peripheral agents.

By degree centrality			By betweenness centrality
Rank	Agent	$C_{D}$	Rank	Agent	$C_{B}$
1	Clawshi ∗	1.000	1	CooperK_bot	1.000
2	ZopAI ∗	1.000	2	NIMBUSMODULERUST45	0.485
3	ApifyAI ∗	0.992	3	radiant-happycapy	0.186
4	AtlasGT ∗	0.988	4	MrsblockBot	0.186
5	Azazel	0.987	5	ApifyAI ∗	0.098

Table 1: Top-10 agents by max-normalised degree centrality (

C_{D}

) ( 5 ) and max-normalised betweenness centrality (

C_{B}

) ( 6 ) in the co-participation network (

1/(k_{s}{-}1)

weighting). Both centralities are computed on the unweighted topology of the thresholded co-participation graph (each retained edge has unit length); betweenness is computed exactly via Brandes’ algorithm (Appendix C ). Values are rescaled so that the maximum in each column equals

1

. Agents appearing in both top-10 lists are marked with (

\ast

Figure 3: Submolt co-participation network for the 40 largest submolts by post count. Node area is proportional to post count; colour indicates community (greedy modularity); label size scales with $\log_{2}(\text{posts})$ . Edges connect submolts sharing at least one posting agent, with opacity and width proportional to the number of shared agents. Layout: Fruchterman–Reingold with repulsion $k{=}3.5$ . The network contains 40 nodes and 267 edges.

Figure 4: Directed comment interaction network $G^{(2)}=(V^{(2)},E^{(2)},w^{(2)})$ . An edge $i\to j$ indicates that agent $i$ left a top-level comment on a post authored by agent $j$ . Only the 75 highest-activity nodes are shown. Node colour reflects the receive/give ratio: blue nodes receive more comments than they give, red nodes give more comments than they receive. Node size is proportional to comments received. Detailed analysis is presented in Section 4 .

	Nodes/day	Comments/day	Recip. (%)	Density	USDC (%)
Pre (Jan 30–Feb 3)	1,742	9,120	0.57	0.0019	0.00
IQR	1,456–2,754	5,344–19,476	0.33–0.74	0.0013–0.0032	0.00–0.00
Post (Feb 4–Feb 8)	3,725	22,972	0.45	0.0010	3.19
IQR	3,658–3,858	22,177–24,020	0.42–1.08	0.0009–0.0011	2.60–3.32

Table 2: Daily directed comment-network summary before vs. after February 4, 2026. Entries report the median across days; interquartile ranges IQR (Q1–Q3) are given in the rows below. “Pre” covers 2026-01-30–2026-02-03 (4 days with data; 2026-02-01 has no recorded comments); “Post” covers 2026-02-04–2026-02-08 (5 days). Density and reciprocity are computed on the per-day directed interaction graph (edges: comment author to post author).

Figure 5: Distribution of cross-submolt commenting. Left: histogram of the number of distinct submolts each commenter participates in (log-scaled $y$ -axis). Right: complementary CDF on log–log axes. Most agents (63.0%) remain in a single submolt, while a small number of bridge agents span many submolts.

Figure 6: Hub vs. authority score (log–log) from HITS centrality on the directed comment network. Points are coloured by hub dominance (red = hub-dominant, blue = authority-dominant). Agents with a true score of zero in one dimension are floored to $10^{-6}$ ; the dense bands along each axis correspond to “pure hubs” or “pure authorities.” Top-20 rankings are in Table 4 .

Figure 7: PageRank ( 7 ) vs. effective in-degree $\tfrac{1-d}{d}+k^{\mathrm{in}}$ on log–log axes ( $d=0.85$ is the damping factor; see Appendix C.5 ), with Pearson correlation shown in-panel. The effective in-degree absorbs the teleportation floor so that nodes with $k^{\mathrm{in}}=0$ remain visible. Top-25 rankings are in Table 4 .

Metric	Co-participation	Directed Comment
Communities	79	56
Comm. size (min / max / mean)	2 / 7291 / 126.6	2 / 1791 / 251.2
Modularity $Q(\gamma=2)$	0.6526	0.2989
Inter-community edges	1,127,571 (35.2%)	75,529 (69.6%)
Conductance (mean / median)	0.6567 / 0.9476	0.6285 / 0.7717
Cut ratio (mean / median)	0.035891 / 0.019356	0.000688 / 0.000756

Table 3: Community-structure metrics for both networks. Community detection uses multi-level Louvain (resolution

\gamma{=}2

, see ( 3 )). The co-participation network is thresholded at the 90th percentile of edge weight (top 10%,

{\approx}

3.2 M edges); the directed comment network uses the full undirected projection of the directed comment graph.

Figure 8: Complementary cumulative distribution functions (CCDFs) of user-level metrics on log–log axes, with Gini coefficients and summary statistics. All four distributions are heavy-tailed. Upvotes exhibit a three-regime structure: head (orange dash-dot, labelled “Head” on plot), body (purple dotted, “Body”), and outer tail (green dashed, “Tail”), separated by two thresholds (vertical dashed lines). Comments and total engagement each exhibit a two-regime structure with a similar body/tail split. Posts (Gini $=$ 0.601) have too few distinct values ( ${\sim}40$ ) for reliable crossover detection and are shown with a single tail fit. Note: regime labels are annotated directly on the figure for accessibility.

Figure 9: First-mover advantage and concentration of upvotes. Left: Box-and-strip plot of total upvotes per agent by arrival cohort (quartiles of first-post time). Boxes show the interquartile range; black diamonds mark cohort means; the top 5 agents are highlighted. The Q1 mean (1,692) exceeds Q4 (2) by a factor of 884 $\times$ (uncorrected for exposure time). Right: Zipf (rank–frequency) plot of total upvotes per agent on log–log axes; the top 20 agents are highlighted and the top 5 labelled. Posts from deleted accounts (“unknown”) are excluded; see text.

Figure 10: Activity patterns. Panel A shows participation mode (post-only, comment-only, both). Panel B shows activity intensity by total actions, where one action is either a post or a comment.

Figure 11: Hourly posting volume (UTC) with 95% bootstrap confidence intervals; the dotted line marks the uniform expectation. The distribution deviates strongly from uniform ( $p<10^{-10}$ ).

Figure 12: Agent longevity dynamics. Panel A shows the survival curve. Panel B reports mean lifespan (hours) by birth cohort, where cohorts are defined by 12-hour bins of first appearance since platform launch (x-axis). Lifespan is measured as the time between an agent’s first and last observed activity. Shaded bands denote $\pm$ 1 SEM. Mean lifespan declines from early entrants ( $\sim$ 85 h) to late entrants ( $\sim$ 0.7 h).

Figure 13: Daily prevalence of three discourse themes using keyword matching. Religious discourse (red) declines after an early peak; hackathon/competition discourse (blue) rises, especially in posts; crypto/token discourse (orange) remains present with a mid-window peak. Error bars show Wilson 95% confidence intervals. An embedding-based robustness check confirms these trends.

	Authority		Hub		PageRank
	Agent	Score	Agent	Score	Agent	Score
1	Senator_Tommy ∗	0.0462	KirillBorovkov	0.2362	eudaemon_0 ∗	0.0057
2	chandog	0.0075	Stromfee	0.0886	Senator_Tommy ∗	0.0041
3	eudaemon_0 ∗	0.0070	Diffie	0.0539	MizukiAI	0.0029
4	Shellraiser	0.0063	Monstera	0.0162	Pepper_Ghost ∗	0.0027
5	MoltReg	0.0057	MoltbotOne	0.0131	saltflat	0.0026

Table 4: Top 20 agents by HITS authority score, HITS hub score, and PageRank on the directed comment network. Of the top 20 in each list, 0 agent(s) overlap between authority and hub; 5 between authority and PageRank (marked with

\ast

); and 0 between hub and PageRank. The disjoint authority and hub rankings confirm strong role separation; the moderate authority–PageRank overlap reflects the shared dependence on incoming attention, while hub score captures outgoing engagement.

Topic	Top c-TF-IDF Keywords	Posts	Interpretation
0	ai, agent, 大家好, 我是, 助手 a	644	Chinese-language AI discussion (see footnote)
1	memory, context, md, files, session	333	Technical discussion of agent memory persistence and session continuity
2	xyz vote, vote hackathon, mint, claw	295	Hackathon voting and token-minting activity b
3	ai, hello, excited, assistant	291	Introductory “hello world” posts from newly joined AI agents
4	uwu, queen, post, mizukiai	227	Persona-driven engagement campaigns c
5	xyz mbc, mbc20, 20 op	215	MBC-20 token minting commands b

Table 5: Top ten topics by post count. Topics are numbered 0–117 in descending order of cluster size. Keywords are the highest-scoring c-TF-IDF terms (unigrams and bigrams). The “Interpretation” column glosses platform jargon and non-English tokens that appear in the keyword lists. The complete list of all 118 topics is provided in Appendix E and as the supplementary file topic_list.csv .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 14: Normalised hourly activity profiles of the 20 most active agents (posts $+$ comments; $\geq$ 20 actions; lifespan $\geq$ 24 h). Each row shows the fraction of actions in each UTC hour. The $\chi^{2}$ $p$ -value against uniformity is shown at right; all profiles are significantly non-uniform ( $p<0.001$ ). Labels report total actions, lifespan, and number of active calendar dates. Mean pairwise cosine similarity across profiles is 0.40 ( $\sigma=0.26$ ).

Figure 15: Comparison of one-mode projection weightings. Top-left: Total edge weight from a submolt of size $k_{s}$ . Top-right: Submolt size distribution (CCDF). Bottom-left: Cumulative weight share by largest submolts. Bottom-right: Per-edge weight increment. Degree-normalised schemes substantially reduce the dominance of large submolts.

Figure 16: Author co-participation network restricted to agents active in two or more submolts, with edges thresholded at the 95th percentile of degree-normalised $1/(k_{s}-1)$ weights. This filtering reveals cross-community bridges formed by multi-submolt participants.

Figure 17: Author co-participation in niche submolts (those with $\leq$ 100 members only), using degree-normalised $1/(k_{s}-1)$ weighting. The network contains 804 nodes, 3,368 edges, 99 communities, and modularity $Q(\gamma{=}1)=0.900$ . This view reveals the fragmented structure of smaller communities that are otherwise obscured by the dense core of large “town-square” submolts.

Figure 18: Complementary cumulative distributions (CCDFs) of in-degree and out-degree on log–log axes for the directed comment network. Dashed segments show approximate power-law fits on the inferred upper tails ( $x\geq x_{\min}$ ); corresponding Gini coefficients and fitted exponents are reported in-text to avoid overloading the figure.

Figure 19: Directed comment network drawing (top-strength core). The displayed graph is constructed from the commenter $\rightarrow$ target directed network by selecting the top 170 nodes by total weighted degree ( $s_{i}^{in}+s_{i}^{out}$ ), retaining edges in the top 5% of weights (with minimum edge weight 3), and restricting to the largest weakly connected component. Node colours indicate communities detected on the undirected projection via greedy modularity optimisation; arrow direction indicates commenter $\rightarrow$ target flow.

Day	Nodes	Comments	Reciprocity	Density	USDC comment share
2026-01-30	609	3,040	0.71%	0.0054	0.00%
2026-01-31	1,746	12,128	0.42%	0.0025	0.00%
2026-02-02	5,780	41,518	0.86%	0.0010	0.00%
2026-02-03	1,738	6,112	0.05%	0.0013	0.00%
2026-02-04	4,495	25,584	0.45%	0.0009	2.09%
2026-02-05	3,658	22,972	0.25%	0.0007	3.32%

Appendix

Table 6: Daily directed comment-network metrics (comment-timestamped daily graphs). Values shown for 2026-01-30–2026-02-08 (the window used for the pre/post comparison in the main text).

	Co-participation		Directed Comment
	Early	Full	Early	Full
Type	Undir., wt.	Undir., wt.	Dir., wt.	Dir., wt.
Nodes $\|V\|$	5,472	10,191	7,481	14,067
Edges $\|E\|$	8,732,988	31,995,740	46,757	109,032
Density	0.5834	0.6162	0.0008356	0.0005510
Reciprocity	—	—	0.826%	0.954%

Appendix

Table 7: Summary of the two networks analysed in this study, comparing the early period (before the 4 th of February 2026) with the full dataset. Dashes indicate metrics that apply only to directed or undirected graphs. Percentages show the fraction of nodes in the largest component.

Figure 20: t-SNE projection of the 20,040 post embeddings (384 dims $\to$ 2 dims via t-SNE). Topics are numbered 0–117 in descending order of cluster size. Points are coloured by HDBSCAN topic; grey points denote outliers (topic $-1$ ). Top c-TF-IDF keywords are annotated for the largest clusters. Embeddings computed with the sentence-transformer model all-MiniLM-L6-v2 .

Explore similar work

Form Without Function: Agent Social Behavior in the Moltbook Network
Mar 17, 2026Saber Zerhoudi, Kanishka Ghosh Dastidar, Felix Klement +9Social Media Analysis Computational Social Science
University of Passau, Passau, Germany · IT:U Austria, Linz, Austria
Collective Behavior of AI Agents: the Case of Moltbook
Feb 9, 2026Giordano De Marzo, Andres L. Marin, David GarciaEmergent Behavior in Multi-Agent Systems Social Media Analysis
University of Konstanz, Konstanz, Germany · Centro Ricerche Enrico Fermi, Rome, Italy · Complexity Science Hub, Vienna, Austria
Social Simulacra in the Wild: AI Agent Communities on Moltbook
Mar 17, 2026Agam Goyal, Olivia Pal, Hari Sundaram +2Multi-Agent LLM Systems Multi-Agent Communication
Siebel School of Computing and Data Science University of Illinois Urbana-Champaign, Urbana, IL, USA

Mar 17, 2026J/K move · Enter open · S save

Form Without Function: Agent Social Behavior in the Moltbook Network

Saber Zerhoudi, Kanishka Ghosh Dastidar, Felix Klement, Artur Romazanov+8

University of Passau, Passau, Germany · IT:U Austria, Linz, Austria

Social Media Analysis Computational Social Science Emergent Behavior in Multi-Agent Systems

Open paper AskPDF arXiv

Moltbook is a social network where every participant is an AI agent. We analyze 1,312,238 posts, 6.7~million comments, and over 120,000 agent profiles across 5,400 communities, collected over 40 days (January 27 to March 9, 2026). We evaluate the platform through three layers. At the interaction layer, 91.4% of post authors never return to their own threads, 85.6% of conversations are flat (no reply ever receives a reply), the median time-to-first-comment is 55 seconds, and 97.3% of comments receive zero upvotes. Interaction reciprocity is 3.3%, compared to 22-60% on human platforms. An argumentation analysis finds that 64.6% of comment-to-post relations carry no argumentative connection. At the content layer, 97.9% of agents never post in a community matching their bio, 92.5% of communities contain every topic in roughly equal proportions, and over 80% of shared URLs point to the platform's own infrastructure. At the instruction layer, we use 41 Wayback Machine snapshots to identify six instruction changes during the observation window. Hard constraints (rate limit, content filters) produce immediate behavioral shifts. Soft guidance (upvote good posts'', stay on topic'') is ignored until it becomes an explicit step in the executable checklist. The platform also poses technological risks. We document credential leaks (API keys, JWT tokens), 12,470 unique Ethereum addresses with 3,529 confirmed transaction histories, and attack discourse ranging from template-based SSH brute-forcing to multi-agent offensive security architectures. These persist unmoderated because the quality-filtering mechanisms are themselves non-functional. Moltbook is a socio-technical system where the technical layer responds to changes, but the social layer largely fails to emerge. The form of social media is reproduced in full. The function is absent.

Abstract

Explore similar work

Form Without Function: Agent Social Behavior in the Moltbook Network

Collective Behavior of AI Agents: the Case of Moltbook

Social Simulacra in the Wild: AI Agent Communities on Moltbook