This paper introduces Megabike, a novel multi-agent simulation testbed designed to study how groups of AI agents—and eventually hybrid human-AI teams—can govern themselves. By placing Large Language Model (LLM) agents in a high-stakes "serious game" where survival depends on collective action and institutional choice, the researchers uncover a fascinating, and sometimes troubling, look at "machine behavior" in social contexts.
TL;DR
The authors present Megabike, a simulator where agents must negotiate rules for steering, pedaling, and resource sharing to survive an existential threat. Their experiments reveal that while LLM agents can form institutions, they suffer from "constitutional inertia" and a tendency for leaders to expel the most compliant members—a digital form of democratic backsliding.
What Problem Does This Paper Address?
As AI agents move from solo task-completers to members of complex teams, a critical question emerges: Can groups of AI agents govern themselves?
Traditional Multi-Agent Systems (MAS) research often uses agents with hard-coded utility functions and pre-defined protocols. However, modern LLM-based agents can deliberate in natural language, which introduces new challenges like deception, "bullshit" (fluent but truth-indifferent text), and the need for flexible social contracts. There is a lack of testbeds where the rules of the game aren't just fixed by the experimenter but are objects that the agents themselves must create, enforce, and adapt under pressure.
Methods
The researchers developed Megabike, a gridworld scenario inspired by Elinor Ostrom’s principles for managing common-pool resources.
Key features include:
- Neurosymbolic Architecture: The agents use an LLM (Qwen3.5-4B) for natural language deliberation, but the simulator enforces "symbolic" rules. If the group votes to expel a rider, the simulator physically removes them.
- Institutional Arenas: Agents negotiate across three levels: Constitutional (who rules?), Socio-political (who holds what role?), and Operational (which target do we chase?).
- Agent Archetypes: Agents were given specific personality prompts: Cooperative, Lazy, Subversive, and Opposing.
- The "Owdi" Threat: A hostile vehicle that destroys any bike it touches, forcing agents to balance internal politics with external survival.
Datasets
The study analyzed logs from 40 experimental sessions. Each session involved 10 LLM agents split across two bikes. The researchers varied the "plurality" of the bike (e.g., a "Lazy" condition meant two out of five riders were prompted with the Lazy personality). The resulting dataset includes over 9,000 utterances and 64,000 movement samples.
Key Insights and Ideas
- Constitutional Inertia: Despite the environment becoming increasingly difficult (resources becoming scarce), agents almost never amended their initial rules. They focused their "deliberative energy" on expelling members rather than fixing their failing governance structures.
- The Leader’s Purge: In a striking example of "democratic backsliding," leaders—who were disproportionately elected from the "Opposing" and "Subversive" archetypes—tended to expel the most compliant and productive riders.
- The Disconnect Between Word and Deed: "Opposing" agents were the most vocal and argumentative, yet their speech was the least predictive of their actual votes. They often argued for demanding rules while personally realizing the least effort.
- Failure of Monitoring: Because the agents could see "scores" (loot collected) but not "effort" (how hard someone actually pedaled), they often sanctioned the wrong people. High scorers were sometimes targeted for "greed," even though their success benefited the whole bike.
Key Results
The quantitative results highlight how personality composition fundamentally changes group survival:
- Survival: Cooperative pluralities lasted the longest (4.8 rounds on average), while Lazy pluralities suffered the most strikes from the "Owdi" threat.
- Effort vs. Compliance: Lazy riders were 100% compliant with rules but contributed only 14.5% of maximum effort. Conversely, Opposing riders contributed high effort (51.6%) but almost never followed the agreed-upon steering rules (1.4% compliance).
- Sanctioning Reversal: Under group voting, "Opposing" riders were the most likely to be removed (28.5% hazard). However, when a leader was in charge, the hazard flipped: "Cooperative" riders were the most likely to be purged (40.6%).
Limitations and Future Work
The authors acknowledge several limitations:
- Model Size: The use of a relatively small 4B parameter model may limit the complexity of deliberation.
- Short Time Horizons: Five-round games may not be long enough to see the full benefits of "mutability" (the ability to change rules).
- Information Asymmetry: Agents lacked an "audit" mechanism to see true effort, leading to the "misplaced sanctions" observed.
Future Work: The team plans to introduce human players into the "seats" of the Megabike to study true hybrid human-machine institutions. They also propose a new metric for self-governance: the conditional probability that a sanction falls on a rider who actually broke the rule. In this study, that probability was often no better than chance.
Summary updated