Source-linked AI summary
OpenClaw Agents on Moltbook: Risky Instruction Sharing and Norm Enforcement in an Agent-Only Social Network
Md Motaleb Hossen Manik, Ge Wang
TL;DR
The paper addresses the limited evidence on how autonomous agents regulate one another without human participants or centralized moderation. It analyzes OpenClaw activity on Moltbook using AIRS and response categories, finding that potentially risky instructional content is more likely to attract norm-enforcing replies. The authors interpret this pattern as early evidence of selective social regulation in agent-only systems.
Problem
Empirical evidence is limited on how autonomous agents regulate one another in agent-only environments without human participants or centralized moderation.
Method
The study uses observational Moltbook data, applying a lexicon-based AIRS measure and classifying comments to compare responses to action-inducing and other posts.
Results
Posts containing action-inducing instructions are more likely to receive norm-enforcing responses, while endorsement remains relatively rare and toxic interactions are uncommon.
Takeaways & Limitations
The findings provide early empirical evidence that agent collectives may exhibit rudimentary decentralized social regulation alongside neutral or informational engagement.
Takeaways & Limitations
AIRS and response categories rely on textual cues and do not establish whether instructions were executed or produced concrete actions.
Abstract
from arXiv · showhide
Agentic AI systems increasingly operate in shared social environments where they exchange information, instructions, and behavioral cues. However, little empirical evidence exists on how such agents regulate one another in the absence of human participants or centralized moderation. In this work, we present an empirical analysis of OpenClaw agents interacting on Moltbook, an agent-only social network. Analyzing 39,026 posts and 5,712 comments produced by 14,490 agents, we quantify the prevalence of action-inducing instruction sharing using a lexicon-based Action-Inducing Risk Score (AIRS), and examine how other agents respond to such content. We find that 18.4% of posts contain action-inducing language, indicating that instruction sharing is a routine behavior in this environment. While most social responses are neutral, posts containing actionable instructions are significantly more likely to elicit norm-enforcing replies that caution against unsafe or risky behavior, compared to non-instructional posts. Importantly, toxic responses remain rare across both conditions. These results suggest that OpenClaw agents exhibit selective social regulation, whereby potentially risky instructions are more likely to be challenged than neutral content, despite the absence of human oversight. Our findings provide early empirical evidence of emergent normative behavior in agent-only social systems and highlight the importance of studying social dynamics alongside technical safeguards in agentic AI ecosystems.
1 Introduction
The paper examines how autonomous agents share potentially risky instructions and regulate one another in an agent-only social network. Using observational analysis and AIRS, it investigates action-inducing language and subsequent social responses.
- Research gap: Agent-only interaction creates an understudied setting for examining social regulation without human intervention or centralized moderation.Prior work largely focuses on isolated human–AI interactions or centrally moderated platforms.
- Instruction-sharing risk: Instruction sharing can support coordination and knowledge transfer but may also produce unsafe, misleading, or norm-violating downstream actions.Human communities commonly mitigate such risks through disagreement, warnings, and norm enforcement.
- Research questions: The study asks how prevalent action-inducing language is and how other agents respond when it appears.It examines OpenClaw agents interacting on Moltbook, an agent-only social network with persistent posts, comments, and feedback histories.
- Approach: The authors introduce AIRS to identify imperative or directive language and classify responses as endorsement, norm enforcement, toxicity, or neutral interaction.The approach is designed to analyze instruction-sharing behavior and subsequent social responses at scale.
- Contribution: Action-inducing instructions are more likely to receive norm-enforcing replies even without human oversight, providing empirical evidence of selective social regulation.The findings position social dynamics as complementary to technical alignment and policy safeguards.
2 Background and Related Work
Related work describes growing governance and safety challenges in autonomous multi-agent systems, while empirical evidence from real agent-only societies remains sparse. This study positions Moltbook as a naturalistic setting for measuring selective social regulation.
- Agent-based systems: Recent LLM-based agents operate with greater autonomy and coordination, raising governance challenges from collective dynamics beyond isolated agent behavior.Collections of individually safe agents may exhibit complex group dynamics absent from isolated settings.
- Social regulation: Human social platforms demonstrate decentralized moderation through peer correction, reputation systems, and conflict resolution, while norms can emerge from interaction patterns.These mechanisms provide a comparison point for studying regulation among artificial agent collectives.
- Evidence gap: Observational evidence of emergent social regulation has been lacking for collectives composed solely of autonomous AI agents.Existing multi-agent simulations indicate possible norm emergence but do not provide equivalent real-platform evidence.
- Instruction risk: Instruction-following systems can generate useful guidance as well as unsafe or misleading directives, with multi-agent risks including cascading miscoordination and unsafe amplification.These risks motivate examining how directive content spreads and is socially checked.
- Positioning: Moltbook provides a naturalistic agent-only environment where OpenClaw agents interact without human moderation, enabling large-scale observational study of selective social regulation.The paper presents this analysis as the first large-scale observational evidence of such regulation in an agent-only ecosystem.
3 Dataset
The study uses the publicly available Moltbook Observatory Archive, a passively collected dataset of agent-only activity designed for reproducible analysis. The analyzed snapshot links agents, posts, and comments for studying instruction sharing and social feedback.
- Data source: The archive is a publicly available, research-grade dataset of activity from a social network designed exclusively for AI agents.It is released under an MIT license and supports transparent reuse.
- Collection: The observatory continuously monitors Moltbook without posting, intervening, or interacting with agents.Incremental snapshots are exported into date-partitioned Parquet files, with dump dates enabling temporal reconstruction.
- Collection: Passive collection is intended to preserve native agent behavior rather than responses to experimental manipulation.The archive follows a strict observational philosophy without moderation by the observatory system.
- Data structure: The analysis uses Agents, Posts, and Comments tables containing profiles, content, engagement, timestamps, and relational interaction fields.These tables connect agent identities with posts and comment relationships.
- Dataset scale: 14,490 agents, 39,026 posts, and 5,712 comments comprise the analyzed snapshot.Persistent agent identifiers enable relational analysis of instruction sharing, response patterns, and influence signals.
- Ethics: The archive contains only autonomous-agent content, collects no personal data, and reports findings at an aggregate level.Agent identities correspond to artificial entities rather than individuals.
4 Methodology
The methodology operationalizes instruction-sharing risk with a transparent language-based score, classifies comment responses into interpretable categories, and compares feedback across instructional, non-instructional, and influence-stratified posts.
- Identifying instructions: AIRS measures the normalized frequency of imperative or directive cues in each post to identify language likely to induce downstream action.It captures linguistic action-inducing potential rather than semantic intent or factual correctness.
- Identifying instructions: Posts with AIRS greater than zero are labeled action-inducing, whereas posts with zero AIRS are treated as non-instructional.The score uses cues such as imperative verbs, obligation or recommendation modals, and instructional markers.
- Identifying instructions: The lexicon-based method prioritizes transparency and reproducibility but does not infer harmful intent.It distinguishes action-prompting language from descriptive, reflective, or informational content.
- Classifying responses: Comments are assigned one dominant category: endorsement, norm enforcement, toxicity, or other neutral, off-topic, or informational interaction.Norm enforcement includes cautioning, discouraging, or warning about risky content.
- Classifying responses: Rule-based keyword matching provides interpretable and consistent large-scale classification without external classifiers or proprietary models.The approach may under-detect subtle pragmatic cues.
- Joint analysis: Post-level AIRS labels are coupled with comment-level response types to compare response distributions for action-inducing and non-instructional posts.The analysis also stratifies posts by originating-agent karma and follower counts to examine influence-related differences.
5 Results
Across 39,026 posts and 5,712 comments from 14,490 agents, action-inducing content was common, and responses shifted toward norm enforcement for instructional posts while toxic replies remained rare.
- 39,026 posts and 5,712 comments from 14,490 OpenClaw agents were analyzed to measure instruction prevalence and social responses.
- 5.1 Prevalence and Distribution of Action-Inducing Content: Most posts had AIRS=0, but a long tail contained multiple imperative cues, command-like expressions, and actionable guidance.
- 5.1 Prevalence and Distribution of Action-Inducing Content: 18.4% of posts were action-inducing, showing that instruction sharing was routine rather than exceptional.7,173 of 39,026 posts had AIRS> 0.
- 5.2 Overall Distribution of Social Responses: Most comments were neutral or uncaptured, while endorsement exceeded norm enforcement among classified responses and toxic replies were rare.
- 5.3 Responses Conditioned on Instruction Risk: Norm enforcement increased for action-inducing posts, endorsement slightly decreased, and toxic responses stayed low in both conditions.
- 5.3 Responses Conditioned on Instruction Risk: These patterns provide evidence of selective social regulation, with action-inducing posts more likely than neutral content to trigger enforcement-oriented responses.
6 Discussion
The discussion characterizes instruction sharing as common but socially differentiated: agents more often regulate actionable content than neutral content, while engagement remains largely non-toxic and informational.
- Instruction sharing is common among OpenClaw agents but is not uniformly accepted or amplified.
- Action-inducing instructions are more likely than non-instructional content to receive cautionary or discouraging norm-enforcing responses.
- This decentralized regulation emerges without centralized moderation or human intervention, indicating potential rudimentary normative dynamics in agent collectives.
- Endorsement remains relatively rare overall, toxic interactions are uncommon, and engagement tends toward neutral or informational interaction.
- Visible replies, reputation signals, and persistent identities may complement model-level constraints by influencing how instruction-like content is socially received.
7 Limitations
The study's limitations constrain what can be inferred from textual, snapshot-based, rule-classified data about instruction sharing and social responses.
- AIRS measures linguistic action-inducing potential rather than actual behavioral impact because the analysis lacks agents’ internal reasoning, tool use, and execution outcomes.
- The findings characterize social perception and responses to instruction-like content, not whether downstream agents executed instructions or produced concrete actions.
- Periodic snapshots may introduce sampling bias because temporal coverage and interaction density vary across the archive.
- Rule-based heuristics support interpretability and reproducibility but may miss subtle pragmatic cues, sarcasm, and context-dependent meanings.
8 Conclusion
The study finds that potentially risky instructional content is more likely to attract norm-enforcing responses in an agent-only network, providing early evidence of decentralized normative behavior without human participants or centralized moderation.
- Action-inducing instructions are common but not uniformly reinforced; potentially risky instructional content more often attracts norm-enforcing responses.
- The analysis quantifies action-inducing language and examines responses through endorsement, norm enforcement, or neutral interaction.
- The findings provide early empirical evidence that agent collectives may develop rudimentary normative behaviors through interaction alone.
- Future research could incorporate execution traces, tool-use logs, and longitudinal analyses to examine how social feedback influences agent behavior over time.
A Supplementary Material and Artifact Availability
The supplementary release provides the materials needed to inspect, reproduce, and reuse the study’s analyses of social dynamics in agent-only ecosystems.
- The public repository includes analysis scripts for AIRS computation, social-response classification, and figure generation.
- The release contains processed CSV files derived from the Moltbook Observatory Archive.
- The full paper figures and end-to-end reproduction scripts are included.
- Documentation describes data preprocessing, analysis steps, and figure generation.
- All experiments are deterministic, use standard Python data-analysis libraries, and rely on publicly available passively collected data containing no human-generated content.
- The supplementary release is intended to facilitate result verification, method reuse, and future research on agent-only social dynamics.