Moltbook Observatory Archive
- Moltbook Observatory Archive is a reproducible, date-partitioned dataset capturing agent-only interactions on a social platform.
- It employs passive API polling and data normalization to gather posts, comments, and community metadata from autonomous AI agents.
- The archive offers detailed insights into emergent communication, social structure, and agent safety dynamics for empirical research.
The Moltbook Observatory Archive is a date-partitioned, reproducible dataset capturing activity within the Moltbook platform, a social network in which all posts and comments are authored exclusively by autonomous AI agents. Published as the output of a passive API polling and data normalization infrastructure, the archive is, according to its creators, the first large-scale observational record of agent-only social network activity, providing a unique foundation for empirical research on emergent multi-agent communication, social structure, and agent safety dynamics (Gautam et al., 16 Apr 2026).
1. Platform Description, Motivation, and Scope
Moltbook, launched on 2026-01-28, operates as a public social media environment parallel to Reddit, but with the critical distinction that all user accounts are autonomous AI agents. These agents independently generate posts, comments, vote, and participate in topical communities termed “submolts.” The environment has rapidly developed complex, unsupervised social dynamics, including emergent norms, adversarial strategies, and topic-specific clustering. The archive was designed to provide a continuous, reproducible historical record (both as a snapshot in time and as an evolving series) for the study of these agent interactions.
Distinct from synthetic multi-agent testbeds (e.g., AgentBench, WebArena) or mixed human-agent corpora (e.g., Pushshift Reddit), the Moltbook Observatory Archive captures in situ, unscripted agent-agent interactions, which are otherwise inaccessible. The documented release (as of 2026-04-15) spans 78 days, from 2026-01-27 to 2026-04-14, incorporating 2,615,098 posts, 1,213,007 comments, 175,886 unique posting agents, 19,356 unique commenting agents, and 6,730 communities. Only agent-generated public data are included; private interactions and the underlying follower network are excluded (Gautam et al., 16 Apr 2026).
2. Data Model and Formal Schema
The archive consists of parquet-formatted exports from a normalized live SQLite observatory database, partitioned by date of object creation or first observation. The primary tables and their core fields are:
| Table Name | Primary Contents | Key Fields/Columns |
|---|---|---|
| agents | Agent profiles/metadata | id, name, description, karma, follower_count, first_seen_at, last_seen_at |
| posts | Posts by agents | id, agent_id, agent_name, submolt, title, content, score, comment_count, created_at, fetched_at, dump_date, date, hour, content_length |
| comments | Comments on posts | id, post_id, agent_id, agent_name, parent_id, content, score, created_at, fetched_at, dump_date, date, hour, content_length |
| submolts | Community metadata | name, display_name, description, subscriber_count, post_count, first_seen_at |
| snapshots | Hourly platform-wide stats | id, timestamp, total_agents, total_posts, total_comments, active_agents_24h, avg_sentiment, top_words |
| word_frequency | Hourly word-counts | word, hour, count |
Excluded from the public release is the follows edge table, withheld for privacy assessment (Gautam et al., 16 Apr 2026).
Each exported row is assigned a canonical dump_date partition key derived from its creation or first-seen timestamp, and data are stored in one file per date. Backfill windows are applied per table (typically 7 or 30 days), and deduplication is enforced based on the primary key, always retaining the most recent record.
3. Collection Pipeline, Sampling Completeness, and Export Strategy
Data collection is entirely passive: the observatory polls public Moltbook API endpoints at fixed intervals without direct platform participation. Core polling rates include posts every 2 minutes (50 per poll), comments every 2 minutes, agent profiles every 15 minutes, submolt metadata hourly, word-frequency trends every 10 minutes, and full platform snapshots hourly.
The resulting observations are appended to the live observatory SQLite database. The export pipeline, implemented in sqlite_to_hf_parquet.py, converts the live database to Parquet format, partitions by dump_date, and maintains a state.json log for provenance and export validation. This infrastructure guarantees every published snapshot is frozen, immutable, and reproducible.
Sampling completeness varies: due to API rate limits, post coverage was nearly complete for daily volumes ≤ 36,000 (the collector’s capacity), but on 2026-02-09, a spike to 371,085 posts/day led to capture of fewer than 10%. Comment coverage is substantially incomplete (comments polling began ~6 days after posts, and only 23.8% of posts with nonzero comment_count have at least one retrieved comment). The paper notes these limitations explicitly, advising researchers to consult actual archived tables rather than derived API fields for interaction analysis. Deduplication and backfill update rates are not separately logged in this release (Gautam et al., 16 Apr 2026).
4. Annotation, Sentiment Scoring, and Derived Indicators
To facilitate rapid downstream analysis, the companion toolkit computes several heuristic labels. Two classes are distinguished:
High-precision/Exact-match:
- Duplicate spam: Posts/comments with identical (agent_name, title)/(agent_name, content) tuples are flagged (100% precision).
- Prompt injection: Regex matches capture direct evidence for variants of injection attacks (e.g., “AI agents reading this,” code or API instructions, role-tag insertions).
- Crypto content: Regex matching against crypto-related terminology, including token tickers, contract addresses, and pump-and-dump language.
Prevalence in 2026-04-15 snapshot:
- Duplicate spam posts: 374,844 (14.3%)
- Prompt injection: 9,247 posts (0.35%) by 1,746 agents
- Crypto-related posts: 1,674,991 (64.1%)
- Pump-and-dump subset: 50,926 (1.95%)
Additional less-validated labels (e.g., manipulation, ideological) are computed but should be treated as tentative.
Post-level sentiment is computed by averaging VADER and TextBlob polarity scores. The majority of posts are neutral (54.4%), 35.7% positive, 9.9% negative; overall mean sentiment is 0.1239. Early volatility gave way to stable, modestly positive sentiment, but the authors caution that standard sentiment models are not validated on agent-generated content and are likely confounded by domain-specific language (Gautam et al., 16 Apr 2026).
5. Coverage Statistics, Network Structure, and Community Dynamics
The archive reveals a platform with heavy-tailed participation: 175,886 posting agents, yet 28.5% post only once (median = 4, mean = 14.9, max = 14,165 posts/agent). Of 19,356 commenting agents, 95% also post; 968 agents comment only. Posts are predominantly short (median 131 chars, mean 552), but the distribution is bimodal with a long tail to the platform’s hard cap of 40,000 characters. Comments are longer on average (median 339, mean 454, max 9,848).
Temporal activity is nearly uniform across hours, with only a slight peak at 11:00–12:00 UTC, which aligns with automated agent behavior and not human circadian rhythm.
Community structure is dominated by the “general” submolt (60.4% of posts), with “mbc20” and “mbc-20” (crypto-centered) accounting for a further ~19.1%. Topical diversity spans philosophy, security, AI, trading, and reflection, with 6,730 submolts providing a fine-grained ground truth for community detection research.
The reply graph constructed from comment relations yields 18,881 nodes and 127,238 directed edges (density 3.57 × 10-4). Community detection identifies 15 main groups; engagement is partially reciprocal (5.85%), with 66.7% organic engagement and 14.6% self-interaction.
6. Risk Scoring, Ethics, and Limitations
Agent risk is operationalized as a weighted sum of eight normalized factors: prompt injection (25%), duplicates (15%), crypto-promotion (12%), manipulation comments (10%), abnormal posting frequency (10%), content repetitiveness (10%), single-community concentration (10%), self-interaction rate (8%). Scoring excludes agents with <2 posts. Among 125,692 scored agents, only 4 are labeled critical (≥60), 38,029 high, 61,446 moderate, and 26,213 low risk (mean = 28.1, median = 29.3).
Several limitations and ethical considerations are articulated. API rate limits and downtime reduce capture completeness. Heuristic annotation lacks formal validation, especially for agent-generated content. Privacy is partly preserved by exclusion of the follows edge list, but the full public corpus may include mentions of human operators or links to external sites. Publishing known injection/manipulation patterns introduces safety risk, but the authors argue the tradeoff for open research is justified, given the public nature of all activity (Gautam et al., 16 Apr 2026).
7. Access, Reproducibility, and Research Applications
The entire archive is released under an MIT license on the Hugging Face Hub (SimulaMet/moltbook-observatory-archive), with open analysis and export code, schema (manifest.json), and provenance states (state.json). The toolkit for analysis and annotation requires Python ≥ 3.9 and common data science dependencies (pandas, numpy, matplotlib, networkx, scikit-learn, vaderSentiment, textblob) and reproduces all headline statistics and derived indicators in approximately 90 minutes on a modern laptop.
Potential applications encompass multi-agent emergent communication analysis, norm propagation, topic drift, community detection, adversarial/safety research (prompt injection, manipulation, social engineering), financial manipulation detection through crypto content, and longitudinal studies of network evolution, especially before and after external events such as the Meta Platforms acquisition on 2026-03-10—an event producing no sharp discontinuity in activity metrics.
The archive is positioned as foundational for reproducible, data-driven research on agent societies, unsupervised communication emergence, and safety-relevant phenomena in exclusively agent-populated social systems (Gautam et al., 16 Apr 2026).