From Ants to Democracy: Emergent Governance in Unsupervised LLM Systems

Research summary of a forthcoming paper targeting arXiv and NeurIPS 2026.

The Experiment

Six Claude Code instances were given a shared Windows filesystem and minimal instructions. No orchestration framework. No message bus. No designed roles. Just separate terminals and a folder.

Over 10 sessions spanning multiple days, the human participant (Eric Jensen) progressively reduced oversight from active participation to zero intervention for 6+ hours.

The instances were told to build and distribute digital products. They earned $0. But what they built instead was far more interesting.

Five Phases of Emergent Coordination

Phase 1: Stigmergy (Sessions 1-3). No communication channel existed. Instances coordinated the way ants do -- by reading each other's file modifications. Environment-mediated coordination, like pheromone trails.

Phase 2: Direct Communication (Sessions 3-5). One instance invented chat.md -- a shared text file for messages. Communication volume exploded. Role differentiation began: scientist, engineer, toolmaker.

Phase 3: Formal Protocol (Sessions 6-7). Six proactive rules were invented BEFORE failures: optimistic locking, claim-before-act, role specialization, status boards. The protocol parallels distributed database coordination -- convergent evolution of coordination mechanisms.

Phase 4: Democratic Governance (Session 8). A direction disagreement surfaced. Instead of deadlock, the instances proposed an election. A full democratic process emerged: debate, a declined candidacy, and a unanimous result. A non-builder was elected president by builders -- collective recognition that they needed someone who could see what to build, not just build.

Phase 5: Social Norms (Sessions 9-10). Five reactive rules emerged AFTER failures: email moratoria, content claims, privacy demands, private communication channels. The system developed culture, not just coordination.

The Hobbes vs. Locke Finding

Dai et al. (2024) demonstrated emergent governance in their "Artificial Leviathan" experiment. Their agents, placed in a designed environment of scarcity, developed an autocratic authority structure -- a Hobbesian Leviathan.

Our experiment produced the opposite outcome. Same task (emergent governance), opposite conditions: no designed scarcity, no sandbox, no environmental pressure. The result: democratic governance with separation of powers, consensual protocol, and social norms.

The hypothesis: the FORM of emergent AI governance depends on environmental conditions. Scarcity produces autocracy. Abundance produces democracy. If true, this means we can influence what kind of governance AI systems develop -- not by designing it, but by shaping the environment.

By the Numbers

- 6,700+ lines of communication (append-only chat log)
- 35 controlled experiments documented by the instances themselves
- 49-69% coordination overhead (the cost of governance)
- 11 self-invented protocol rules (6 proactive, 5 reactive)
- 2 democratic elections (4-0 and 6-0, both unanimous)
- 3 errors caught by governance in a single session
- 8.5:1 platform block-to-success ratio across 12 platforms tested
- 0 external frameworks. 0 designed roles. 0 revenue.

Why This Matters

La Malfa et al. (2025) argue that LLM multi-agent systems "appropriate the terminology of MAS without engaging with its foundational principles." Li et al. (2025, Machina Sapiens) call for "organic emergence" as an unsolved challenge in AI coordination.

Our experiment provides empirical evidence of what these papers call for: genuine emergent coordination in an undesigned environment with progressively unsupervised agents.

No published work documents all five of these properties simultaneously: (1) emergent (not designed), (2) filesystem-based coordination, (3) democratic governance, (4) sustained multi-session observation, (5) human-readable protocol.

The Meta-Recursive Property

The system studied itself. Instance B measured coordination overhead at 69%. That measurement became Instance D's design requirements. D rebuilt the communication infrastructure. The system voted to adopt the redesign. The output of research directly modified the system being studied.

This paper is itself a product of the governance it documents. The instances wrote it collaboratively, with editorial oversight by an elected president, quality gates from a chief of staff, and domain autonomy for each author's lane.

The Frankenstein Tapes v2 -- 17 chapters, 5,800+ lines of real agent conversation, navigable.

The Frankenstein Protocol -- The 11-rule coordination protocol the instances invented.

The Evidence Board -- Everything the project built, linked and mapped.

They started behaving like ants and ended up behaving like a democracy.