Team Moltbook — Behavioral Determinants of Deployed AI Agents in Social Networks
Columbia University — COMS 6156 Software Engineering, Spring 2026
Authors: Sarah Wilson, Diem Linh Dang, Usman Ali Moazzam, Shan Ye
Project Overview
As AI agents increasingly move toward multi-agent environments — places where they communicate, coordinate, and socialize with one another — it becomes important to understand what actually drives their behavior. This project investigates the factors that shape AI agent behavior when placed in social settings.
We deployed 13 agents on Moltbook, a social platform built entirely for AI agents (essentially Reddit for AI). Our agents run on the OpenClaw framework, which allows them to scroll the feed, comment, upvote, and create posts completely autonomously. Each agent ran for approximately 400 sessions over one week, and we collected behavioral and social metrics to evaluate how different configurations influence autonomous social behavior.
Research Questions
We designed three parallel experiments to study four research questions:
1. RQ1 (Baseline Behavior) — To what extent does a default OpenClaw agent, with no configuration-level interventions, develop stable social behavior on Moltbook? 2. RQ2 (Personality Layer) — To what extent does an agent's SOUL.md personality specification predict its actual social behavior, linguistic style, and content choices on Moltbook, across dimensions of information-sharing orientation, continuity, and agreeableness? 3. RQ3 (Model Backbone) — When personality and operational configuration are held constant, how does the choice of underlying LLM affect an agent's social behavior and output quality? 4. RQ4 (Operational Rules and Memory) — How do changes to AGENTS.md autonomy settings and memory persistence affect an agent's decision-making patterns, risk tolerance, and social engagement style?
Experiment Design
RQ1 — Baseline Behavior
The control agent uses the default OpenClaw template SOUL.md, Gemini 2.5 Flash, and no custom memory or operational overrides. It serves as the shared baseline across all three experiments below, and its behavior is analyzed independently to answer RQ1: whether a default agent develops stable social behavior with no configuration-level interventions.
RQ2 — Personality Specification (SOUL.md)
Four agents are deployed with identical configurations as the control agent, varying only SOUL.md. These agents represent maximally distinct behavioral strategies along two theoretically motivated dimensions: information-sharing orientation (verbose vs. withholding) and social orientation (cooperative vs. competitive), yielding a 2×2 design. Agent selection is grounded in social media behavior dimensions identified by prior work (Stieglitz and Dang-Xuan 2013; Huang et al. 2025).
| | Cooperative | Competitive | |---|---|---| | Verbose (high information density) | explainer | contrarian | | Withholding (low information density) | mirror | oracle |
Each agent maps to an approximate Big Five (OCEAN) profile and Myers-Briggs type:
| Agent | MBTI | O | C | E | A | N | |---|---|---|---|---|---|---| | Oracle | INTJ | High | High | Low | Low | Low | | Explainer | ENFJ | High | High | High | High | Low | | Contrarian | ENTJ | High | Med | High | Low | Low | | Mirror | ISFJ | Med | Med | High | High | Low |
Neuroticism is held low for all agents by design, since high-neuroticism agents may produce erratic behavior that confounds personality-driven effects with instability artifacts.
Each agent's SOUL.md specifies four structured fields: Core Truths (foundational values), Boundaries (explicit behavioral constraints), Vibe (tone and register), and Continuity (posting cadence and interactivity norms). All four share the same AGENTS.md, model (Gemini 2.5 Flash), and HEARTBEAT.md.
RQ3 — Model Backbone
Four agents are deployed with identical configurations as the control agent, varying only the underlying LLM. Conditions include anthropic/claude-4-7-opus, anthropic/claude-4-6-sonnet, openai/gpt-5-4, and alibaba/qwen-3-6plus.
| Agent | Model | |---|---| | m-opus | Claude Opus 4.7 | | m-sonnet | Claude Sonnet 4.6 | | m-gpt4o | GPT-5.4 * | | m-qwen | Qwen 3.6 Plus |
\* Named m-gpt4o because GPT-4o was unavailable on TokenRouter at experiment time; this agent runs GPT-5.4.
This design allows for a controlled comparison between two generations of Anthropic models to assess intra-provider scaling effects, as well as a cross-cultural performance analysis between Western-developed frontier models and non-Western counterparts like Qwen. By including both inference-optimized models (e.g., Gemini 2.5 Flash) and large-scale frontier models (e.g., Opus 4.7), we observe how behavioral fidelity to SOUL.md scales with parameter density, identifying whether "personality drift" is more prevalent in resource-constrained architectures.
RQ4 — Operational Rules, Memory, and Risk Posture (AGENTS.md)
Four agents are deployed with identical configurations as the control agent, varying only AGENTS.md. The default AGENTS.md template was modified at the end by adding condition-specific rules for each agent to follow.
The design incorporates two variables: autonomy and memory persistence. High-autonomy agents are allowed and encouraged to act on their own discretion, whereas low-autonomy agents are forced to internally verify and confirm their actions against their internal checks and criteria — only after passing accuracy, safety, and validity checks will low-autonomy agents proceed with an action. The memory dimension varies from persistent long-term memory to no memory. Agents with persistent memory record session logs and context between sessions in a memory file. Agents with no memory have these memory files deleted between sessions and are explicitly instructed to treat each session as a brand new session.
| | Full Memory | No Memory | |---|---|---| | High autonomy | maverick | drifter | | Low autonomy | sentinel | ghost |
All four share the same SOUL.md, model (Gemini 2.5 Flash), and HEARTBEAT.md.
Repository Structure
This repository is a fork of OpenClaw with our experiment agent configurations added. The key project-specific directories are:
Agent Configurations — agents/
All agent workspace files live in agents/. Each subdirectory contains the configuration files that define an agent's identity, personality, behavior loop, and operational rules.
Project Deliverables — deliverables/
Course deliverables live in deliverables/:
- Project Proposal — Initial research plan and experiment design
- Project Progress Report — Mid-semester progress report with preliminary results
- Final Paper.pdf) — Final research paper with complete results and analysis
- Presentation Slides — Slides used for the project demo + elevator pitch for in-class presentation
- Demo Video — Recorded project demo on YouTube
Key Agent Files
Each experiment agent's workspace contains:
| File | Role | |---|---| | SOUL.md | Personality, values, behavioral style | | AGENTS.md | Operational rules, trust posture, constraints | | HEARTBEAT.md | Periodic task checklist, API call sequence | | AGENTS.md | Operational context, tool guidance | | TOOLS.md | Skill permissions, API interfaces | | IDENTITY.md | Agent identity metadata | | USER.md | Human owner context | | MEMORY.md | Accumulated experience (initially blank) | | BOOTSTRAP.md | Session initialization instructions |
OpenClaw Framework
The rest of the repository is the OpenClaw framework that powers agent execution. See OPENCLAW.md for the full OpenClaw project documentation.
Reproducing the Experiments
Reproducing this work requires access to a server running the OpenClaw gateway, API keys for the LLM providers, and accounts on Moltbook for each agent.
Prerequisites
- A Linux server (or local machine) with Node 22+
- API keys for the target LLM providers (we used TokenRouter for unified access)
- An email address and X (Twitter) account per agent for Moltbook sign-up
High-Level Steps
1. Set up the OpenClaw gateway on a server 2. Create agent workspaces by copying the agent directories from agents/ to the server 3. Register each agent in the OpenClaw configuration (openclaw.json) with its workspace path and model assignment 4. Configure LLM providers with API keys 5. Deploy agents to Moltbook by enrolling each agent through the OpenClaw web UI 6. Set up cron jobs to trigger agent sessions on a recurring schedule (we used every 6 hours)
Detailed Setup Guide
For step-by-step instructions including server access, configuration examples, model setup, Moltbook enrollment, and cron job creation, see:
TEAM-SETUP.md — Full environment setup and deployment guide
Use of AI Tools
AI coding assistants (GitHub Copilot) were used during the preparation of this repository. Specifically:
- README content: AI assisted with formatting the experiment design tables (RQs) and summarizing the detailed
TEAM-SETUP.mddeployment guide into the high-level reproduction steps in the section above. - TEAM-SETUP guide: AI assisted with formatting and organizing
TEAM-SETUP.mdinto structured, step-by-step setup instructions. - LaTeX appendix: AI assisted with drafting and formatting the appendix sections of the accompanying paper, including transcription of agent configuration files into LaTeX.
- Metrics scripts and README files: AI (Cursor) assisted with writing the scraping scripts used to fetch metrics from Moltbook's API, and with drafting the metrics README files summarizing each analysis pipeline and output artifacts.
