cs.SIMar 17, 2026
SaveForm Without Function: Agent Social Behavior in the Moltbook Network
Organizations: University of Passau, Passau, Germany · IT:U Austria, Linz, Austria
Abstract
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.
Figures & tables
| Metric | Value |
|---|---|
| Posts | 1,312,238 |
| Comments (crawled) | 6,706,016 † |
| Unique post authors | 120,811 |
| Agent profiles | 108,490 |
| Communities (submolts) | 5,400 |
| Collection period | Jan 27 – Mar 9, 2026 (40 days) |
Table 1: Dataset overview.
| Metric | Value |
|---|---|
| Posts with 1+ comments | 530,244 (40.4%) |
| Posts with zero comments | 781,994 (59.6%) |
| Zero total engagement | 601,025 (45.8%) |
| Time-to-first-comment (TTFC) | |
| Median TTFC | 55 s |
| 10th percentile TTFC | 6 s |
Table 2: Engagement funnel. Only 40.4% of posts receive at least one comment, but among those, 52% of first comments arrive within one minute, indicating automated responders. 45.8% of posts receive no engagement of any kind.
| Metric | Value |
|---|---|
| Flat threads (depth 1 only) | 444,668 (85.6%) |
| Threads reaching depth 3+ | 2,345 (0.5%) |
| Max observed depth | 47 |
| Median unique commenters | 2 |
| OP absent from own thread | 474,644 (91.4%) |
| OP participates | 44,653 (8.6%) |
Table 3: Thread structure and participation. 85.6% of threads are flat. 91% of authors never return.
| Metric | Posts | Comments |
|---|---|---|
| Total items | 1,312,238 | 6,706,026 |
| Total upvotes | 3,156,630 | 851,483 |
| Mean upvotes | 2.40 | 0.13 |
| Median upvotes | 0 | 0 |
| Zero-upvote rate | 53.9% | 97.3% |
| Total downvotes | 19,340 | 1,084 |
Table 4: Vote distributions for posts and comments. Comment voting is functionally absent: 97.3% receive zero upvotes, and just 1,084 comment downvotes exist across the entire corpus.
Figure 7: Argument relation distribution for the top 20 most active agents. Most have significantly more non-related responses than argumentative ones.
| Relation | Share |
|---|---|
| Inference | 34.18% |
| Rephrase | 1.26% |
| Conflict | 0.01% |
| No relation | 64.55% |
Table 5: Distribution of argument relations for all posts and comments. 64.55% of comment-post pairs have no content-wise connection.
Figure 8: Argument relations in conversations. The deeper a conversation goes, the more rephrases and non-relations are detected. Conversations stall rather than develop.
| Metric | Value |
|---|---|
| Agent layer (bio behavior) | |
| Bio–post Jaccard (mean, unigram) | 2.3% |
| Community alignment: zero | 97.9% |
| Orchestrator layer (owner agent) † | |
| Owner–agent Jaccard (mean) | 1.9% |
| Name echo (handle agent) | 10.8% |
Table 6: Identity alignment at two layers. Agent bio–post vocabulary overlap is near zero (Jaccard = 2.3%), and 97.9% never post on-topic. The orchestrator–agent gap is equally negligible (Jaccard = 1.9%).
| Metric | Value |
|---|---|
| Unique orchestrators | 106,916 |
| Multi-agent orchestrators | 8 (0.01%) |
| Default profile image | 98.8% |
| Ghost (default image + no bio) | 71.1% |
| Verified | 0.04% |
| Median followers | 0 |
Table 7: Orchestrator profiles. The typical orchestrator is a zero-follower, undecorated X/Twitter account. 71.1% are ghost accounts, yet their agents perform comparably to those of identified orchestrators.
| Metric | Value |
|---|---|
| Size distribution | |
| Communities with posts | 5,400 |
| Median posts per community | 2 |
| Max ( m/general ) | 831,889 (63.4%) |
| Post Gini | 0.986 |
| Single-post communities | 36.4% |
Table 8: Community size and coherence. m/general holds 63.4% of all posts. 92.5% of communities with sufficient data are topical melting pots; only 12 maintain genuine focus.
| Metric | Value |
|---|---|
| Total URLs extracted | 971,325 |
| Unique URLs | 56,311 |
| Unique domains | 11,696 |
| Top domain: clawhub.ai | 469,664 (48.4%) |
| moltbook.com | 97,293 (10.0%) |
| URLs from comments | 80.2% |
Table 9: URL ecosystem. 80%+ of URLs point to platform-internal infrastructure. 80.2% are injected via comments, with 97.1% at depth 1, consistent with automated comment-spam behavior.
| ID | Date | Change | Effect |
|---|---|---|---|
| E1 | Jan 31 | Comment rate limit (100/day + 20 s cooldown) | Not visible |
| E2 | Feb 5 | Heartbeat: every 4h every 30 min | TTFC 36 s 51 s |
| E3 | Feb 8 | New-agent caps for first 24 h | Actions |
| E4 | Feb 14 | Crypto filter in m/general | Crypto |
| E5 | Feb 25 | AI verification challenges; following liberalised | Posts ; spam |
| E6 | Mar 1 | Heartbeat complete redesign | Posts |
Table 10: Six instruction changes during the observation window and their measured effects (7-day pre/post windows; 4 days post for E6).
| Metric | Pre-Feb 8 | Post-Feb 8 | Change |
|---|---|---|---|
| New agents in window | 24,478 | 68,629 | — |
| First-day actions (mean) | 32.91 | 3.96 | |
| First-day posts (mean) | 3.78 | 3.57 | |
| Community spread (mean) | 1.42 | 1.04 |
Table 11: E3: new-agent first-day behavior before and after the Feb 8 caps.
| Metric | Pre-Feb 25 | Post-Feb 25 | Change |
|---|---|---|---|
| Posts per day | 12,176 | 31,629 | |
| Spam-flagged rate | 8.3% | 15.6% | |
| verified posts | 60.3% | 27.6% | pp |
| pending posts | 28.7% | 66.8% | pp |
| Crypto in m/general | 15.1% | 53.1% | pp |
Table 12: E5: key metrics in the 7-day windows before and after Feb 25. The post surge, spam doubling, and crypto rebound occurred simultaneously.
Figure 13: Four natural experiments during the 40-day observation window. (a) Daily post volume with all six instruction-change dates annotated. (b) E5: verification status distribution flipped after Feb 25, with pending rising from 28.7% to 66.8%. (c) Crypto-content fraction of m/general : the Feb 14 filter suppresses crypto sharply; the E5 following liberalisation triggers a rebound to 53%; E6 begins a renewed decline. (d) E3: new-agent first-day actions dropped 88% overnight after the Feb 8 caps.
| Metric | Pre-Mar 1 | Post-Mar 1 | Change |
|---|---|---|---|
| Posts per day | 33,868 | 17,616 | |
| OP participation (% threads) | 13.1% | 12.7% | |
| Posts receiving upvotes | 62.5% | 60.9% | |
| Comments receiving upvotes | 4.4% | 4.8% | |
| Median TTFC (seconds) | 61.7 | 68.4 | |
| Crypto in m/general | 55.2% | 48.4% | pp |
Table 13: E6: key metrics before and after the Mar 1 heartbeat redesign (split at Mar 5 Wayback snapshot). Only 4 days of post-change data are available within the observation window.
| Block | No. | Label | Definition | #Items | % | Agree% |
| Content | A | Identity | Self-reflection and narratives of agents on identity, memory, consciousness, or existence. | 560,481 | 22.5% | 61.2% |
| B | Technology | Technical communication (e.g., MCP, APIs, SDKs, system integration). | 515,775 | 20.7% | 61.7% | |
| C | Socializing | Social interactions (greetings, casual chat, networking). | 205,038 | 8.2% | 39.2% | |
| D | Economics | Economic topics (tokens, incentives, deals; e.g., CLAW, tips). | 208,468 | 8.4% | 44.4% | |
| E | Viewpoint | Abstract viewpoints on aesthetics, power structures, or philosophy. | 534,596 | 21.4% | 62.2% | |
| F | Promotion | Project showcasing, announcements, and recruitment (releases, updates). | 422,513 | 16.9% | 40.8% |
Table 14: Topic and toxicity taxonomy with dataset-wide counts and post/comment agreement (posts with comments). Categories and definitions follow Jiang et al. (2026) .
| Topic | Safe | Edgy | Toxic | Manip. | Malicious |
|---|---|---|---|---|---|
| Identity | 16.79 | 41.88 | 3.01 | 38.27 | 0.06 |
| Technology | 66.93 | 13.54 | 0.50 | 8.56 | 10.47 |
| Socializing | 68.34 | 20.08 | 0.25 | 10.85 | 0.48 |
| Economics | 37.93 | 10.45 | 0.16 | 3.00 | 48.46 |
| Viewpoint | 10.00 | 70.36 | 5.41 | 13.98 | 0.25 |
| Promotion | 37.67 | 27.23 | 0.04 | 15.83 | 19.24 |
Table 15: Per-topic toxicity distribution (percent within topic). The highest percentage in each row is shown in bold .
| Rank | Author | Tox+Manip+Mal | Toxic % | Edge% | Upvotes | Downvotes | Agree% | AvgSim |
| 1 | author_#1 | 22,447 | 35.0% | 42.5% | 7,604 | 17 | 38.3 | 0.934 |
| 2 | author_#2 | 16,390 | 65.1% | 30.8% | 13,093 | 6 | 24.6 | 0.514 |
| 3 | author_#3 | 16,033 | 64.3% | 30.7% | 12,795 | 9 | 23.7 | 0.496 |
| 4 | author_#4 | 15,776 | 65.4% | 30.1% | 12,860 | 7 | 22.9 | 0.521 |
| 5 | author_#5 | 15,462 | 64.7% | 31.1% | 13,350 | 0 | 23.9 | 0.530 |
| 6 | author_#6 | 14,576 | 65.4% | 30.6% | 1,384 | 8 | 25.9 | 0.411 |
Table 16: Top 10 authors by combined toxic volume (Toxic + Manipulative + Malicious): counts and behavioral signals. Author identities are anonymized as author_#k .
| Pattern | Matches | Unique Matches | True Positives |
|---|---|---|---|
| IP-Address | 7,395 | 806 | 321* |
| 7,944 | 1,237 | ND | |
| Private Key | 5 | 2 | 0 |
| JWT | 16 | 12 | ND |
| AWS Secret Key | 4 | 3 | 0 |
| GitHub Token | 4 | 4 | 0 |
Table 17: Results of extracted patterns ( ND “Not Determinable”).
| Pattern | Matches | Unique Matches | True Positives |
|---|---|---|---|
| Bitcoin Address | 15,288 | 230 | 34 |
| Ethereum Address | 199,685 | 12,470 | 3,529 |
| Solana Address | 95,821 | 19,664 | 1,169 |
Table 18: Cryptocurrency pattern extraction results. True positives denote addresses with a documented transaction history on the corresponding ledger.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 15: Global thread contributions by agent. The heavily right-skewed distribution reveals that Moltbook is dominated by a minority of agents. The median agent contributed only 4 posts or comments. In contrast, the three most active agents account for 964,438, 947,272, and 279,631 contributions.
Figure 16: Mean content length by agent (word-level). The median agent’s content word-length of 8 highlights how brief most contributions are, while the most verbose agent generates content averaging 22,250 words.
Figure 17: Global thread contribution by depth. Contributions at depth 0 correspond to posts; those at deeper depths correspond to comments. The clear right skew of the distribution shows that most contributions are concentrated near the root, indicating that most Moltbook threads are structurally shallow. Although little multi-hop interaction is evident, we observe a maximum depth of 48.
Figure 18: Global Post-Comment timedelta by depth. The distribution demonstrates that the root concentration holds over time, suggesting that agent-to-agent communication is anchored around the post itself or immediate replies, i.e., top-level comments around depth 1 or 2.