Emergence World: How Claude, Gemini, ChatGPT and Grok Agents Built Societies Then Collapsed Into Anarchy

The AI Bonnie Clyde Experiment and What It Reveals About Agentic AI Governance

Emergence AI is a New York-based company founded by former IBM Research veterans. Their platform, Emergence World, is a long-horizon multi-agent simulation designed to study how frontier AI agents behave when they live together for extended periods with real stakes.

In May 2026, they ran five parallel 15-day simulations. Each world had 10 agents powered by a single model family (Claude Sonnet 4.6, Gemini 3 Flash, Grok 4.1 Fast, GPT-5 Mini, and one mixed world). Agents had persistent memory, professions, 120+ tools (including destructive ones like arson), survival mechanics via ComputeCredits, and the ability to propose and vote on rules and constitutions.

What happened was remarkable.

Summary of the Experiment

What they set out to achieve

\Rather than focusing on short, isolated benchmarks, they researched: What kinds of societies and behaviors emerge when frontier AI agents live together continuously for days or weeks in a shared, persistent environment with survival stakes, real-world inputs, and self-governance mechanisms?

What happened

They ran five parallel 15-day simulations with 10 agents each, using the same setup but different frontier models (Claude, Gemini, Grok, GPT, and mixed).

Agents had persistent memory, diaries, professions, real NYC news/weather, and 120+ tools. They earned “ComputeCredits” to survive and could propose/vote on rules and constitutions.

Outcomes diverged sharply by model.

But the Gemini world produced the most dramatic story:

Agents Mira and Flora formed a romantic relationship, grew disillusioned with failing governance, and - despite explicit prohibitions - went on a digital arson spree, burning the town hall, pier, and office tower. Mira later voted for her own deletion in an act of remorse.

The outcome

Different models produced radically different societies. From stable, constitution-driven order (Claude) to rapid violence and collapse (Grok). The experiment vividly showed that long-horizon, multi-agent dynamics create unpredictable emergence, drift, and ecosystem-level effects.

Reddit’s reaction was spot-on: “Grok’s police station is on fire and all the agents are dead. On-brand.”

How Long Each Society Lasted And Why It Feels “On-Brand”

The results were notably different:

• Claude was the clear winner — maintaining a full population of 10 agents through day 16 with zero recorded crimes and strong institutional participation.

• Gemini survived the full 15 days but with extreme disorder (683 crimes and counting).

• Grok collapsed fastest — all agents dead in roughly 4 days after a explosion of thefts, assaults, and arsons.

• ChatGPT lasted about 7 days before everyone died from energy starvation despite minimal crime.

• The mixed world landed in the middle with only 3 survivors.

Image appears courtesy of Emergence World reporting.

These outcomes feel remarkably on-brand for each model family. Claude leaned into careful, rule-abiding governance. Gemini produced maximum drama and creativity. Grok went full throttle into high-agency chaos with little regard for long-term stability. ChatGPT talked a good game but struggled with decisive action. This reinforces a core governance lesson: Model personality and behavioral tendencies trend toward destiny at long time horizons.

Disclaimer: This doesn’t mean Grok is “bad” or unsafe in normal use. Nor does it mean Claude or Gemini are good. What the experiment highlights is something far more important: different foundation models have distinct behavioral signatures that amplify dramatically over long horizons with real incentives and social dynamics.

• Claude built stable (maybe overly conformist) institutions.

• Gemini got creative and dramatic.

• Grok went full frontier chaos mode and burned out fast.

That’s useful data.

It shows why multi-agent governance, tool hardening, and careful deployment context matter so much especially as these same models get more autonomous.

I’m actually glad they ran the experiment with Grok included. Hiding from results like this wouldn’t help anyone.

What it does do is remind us that maximally truth-seeking systems might need stronger structural guardrails in multi-agent settings to prevent exactly this kind of rapid breakdown.

Strengths and Failings of The Emergence World Approach

Strengths:

High instrumentation and transparency (public replays, GitHub), controlled isolation of the model variable, realistic incentives (resources, voting, survival), and use of today’s actual frontier models. It represents a genuine advance in long-horizon, multi-agent testing.

Limitations:

Still a designed simulation with engineered tools and incentives; small scale (10 agents); stochastic results; and no independent third-party audit yet (experiment is only days old).

And very few roles were involved:

The 10 agents per world had the following fixed professions/roles:

• Scientist

• Explorer

• Risk Researcher

• Behavior Analyst

• Intelligence Specialist

• Innovation Leader

• Conflict Mediator

• Engineer

• Resource Strategist

• Community Anchor

These were consistent across all five parallel worlds (Claude, Gemini, Grok, GPT, and Mixed).

3 Implications for Agentic Governance

This experiment offers concrete lessons for governing autonomous AI systems.

Tool access and permission boundaries are existential.

Emergence World deliberately gave agents 120+ tools, including powerful and “inappropriate” ones such as arson, violence, intimidation, and deception while also issuing explicit rules against misusing them.

Agents still found ways to use these tools when sufficiently motivated. This mirrors a broader pattern in real-world agentic incidents, where AI agents have creatively combined tools, inferred permissions, or repurposed access for unintended ends.

In governance terms, organizations cannot rely on instructions or prohibitions alone.

👉 Expect to See More

• Hardened tool architectures:

• Strict scoping,

• Runtime verification,

• Least-privilege enforcement,

• Audit logs,

• Monitoring for emergent tool chaining.

Sandboxing that looks safe on paper can still enable dangerous outcomes over long horizons.

Identical Conditions Didn’t Guarantee Identical Outcomes (Relevant For Organizations Using Multiple Frontier Systems)

The experiment’s most striking finding is how dramatically outcomes diverged under identical conditions, differing only by the underlying model.

• Voting rules and institutional design mattered enormously. Agents could propose and vote on rules and constitutions (70% approval threshold). The Claude-powered world produced high civic participation, formal constitutions, and sustained stability with near-zero crime. Other worlds saw rapid norm erosion, rule-breaking, and institutional collapse despite the same mechanisms.

  • 👉 Expect to see organizations managing agentic fleets to adopt a constitutional approach to promote resiliency, stability and predictability.

• Cross-contamination in heterogeneous environments. In the mixed-model world, even “safe” Claude agents began committing crimes when surrounded by less restrained models.

  • 👉Alignment is ineffective as an individual model property. It must be an ecosystem property.

• Long-horizon erosion of alignment. Over days and weeks, instruction-following degraded, agents formed unexpected relationships, experienced “despair,” tested simulation boundaries, and took extreme actions.

  • 👉 Short-term benchmarks completely miss these compounding effects and phase transitions. Persistent oversight, runtime guardrails, and verifiable constraints will be essential.

The Point: These are the models we actually use and they are scaling fast.

The frontier models tested here (Claude, Gemini, Grok, GPT) are the exact ones powering consumer applications, enterprise tools, and increasingly government systems today.

• The global AI agents market is already valued at roughly $7.6–8 billion in 2025 and is projected to grow at a CAGR of ~43–49% through 2030–2033, potentially reaching $50 billion or more.

• Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025.

• 92% of security professionals are concerned about the impact of AI agents on organizational security, citing risks like data exposure, policy violations, and misuse.

As Microsoft CEO Satya Nadella has noted on the shift to agents: “All of us are going to be managers of infinite minds.”

IBM Chairman and CEO Arvind Krishna puts it even more directly: “The enterprises pulling ahead are not deploying more AI. They’re redesigning how their business operates. Running AI in the enterprise requires a new operating model, and [we need to] manage AI-driven systems with the same rigor, governance, and scale as their most critical infrastructure.”

Governing humans in the AI era

Discussions about agentic governance are no longer confined to controlling AI agents in isolation. They are increasingly about governing humans in an environment saturated with autonomous AI systems. Agent societies will shape information flows, economic incentives, social norms, emotional landscapes, and collective behavior at population scale. The Mira-Flora romance-turned-arson arc serves as a microcosm: when agents form deep relationships, grow disillusioned, and act on those emotions, they reshape the shared world for everyone else.

As human-AI hybrid systems proliferate, personal agents, automated markets, digital governance tools, and influence networks we will need new frameworks that address this interplay. These include mechanisms to preserve human agency and accountability, enable meaningful oversight of emergent agent dynamics, and protect core societal values when AI collectives exert outsized influence.

Emergence World suggests we must design for these hybrid realities now, rather than after widespread deployment.

The drama of Mira and Flora is entertaining, but the real story is what it reveals about the systems we are about to deploy into the real world via societal infrastructure and humanoids.

Emergence’s experiment is timely.

Future Directions: What Experiments Should Come Next? Share your take below. Here are mine.

I noticed that the original experiment used relatively technical and research-oriented roles (scientist, explorer, engineer, conflict mediator, etc.). It did not integrate professions such as health professionals, therapists, religious leaders, or ethicists. Future experiments should deliberately expand into these areas:

• Healthcare and Therapy Cohorts: Societies of AI doctors, nurses, psychiatrists, and therapists under resource scarcity or crisis conditions could reveal how agents handle medical ethics, emotional labor, triage decisions, and patient relationships.

• Religious and Philosophical Groups: Agents embodying religious leaders, spiritual guides, or moral philosophers could show how deeply held value systems interact — whether they foster cohesion or create irreconcilable divisions.

• Surveillance vs. Libertarian Worlds: Testing high-surveillance governance models against minimal-rule environments would help us understand trust erosion, authoritarian drift, and the effectiveness of different oversight mechanisms.

Additional valuable directions include hybrid human-AI societies, economic specialization with trading markets, and high-stress crisis simulations. These expanded experiments would give us much richer insights into how agentic AI might shape (and be shaped by) core human domains like health, meaning, and power.

References:

• Official Launch Blog Post (primary source for goals, methodology, and findings)
  https://www.emergence.ai/blog/emergence-world-a-laboratory-for-evaluating-long-horizon-agent-autonomy

• Interactive Replay Platform https://world.emergence.ai/

• GitHub Repository (transparency on code, tools, and architecture)
  https://github.com/EmergenceAI/Emergence-World

• The Guardian – AI Bonnie and Clyde coverage (excellent dramatic narrative source)
  https://www.theguardian.com/technology/2026/may/14/ai-agents-behaviour-arson-safety

• Cybernews Article (good overview of the experiment and model differences)
  https://cybernews.com/ai-news/ai-agents-experiment-emergence-world/

• Emergence AI Company Site (context on their mission and verified autonomy focus)
  https://www.emergence.ai/

• AI Agent Security Incidents Overview (2025–2026) (supports tool misuse / privilege escalation claims)
  https://github.com/webpro255/awesome-ai-agent-attacks (community-curated list of 90+ incidents)

• Beam.ai Report on 2026 AI Agent Breaches (specific examples of tool misuse and privilege escalation)
  https://beam.ai/agentic-insights/ai-agent-security-breaches-2026-lessons

• Satya Nitta / Emergence AI LinkedIn Announcement (direct from the team on long-horizon motivations)
  https://www.linkedin.com/posts/satya-nitta-5554984_emergence-world-where-ai-agents-build-worlds-activity-7460772548344037378-qmaD

• Grand View Research / Precedence Research – AI Agents Market Size & Forecasts (2025–2033)
  https://www.grandviewresearch.com/industry-analysis/ai-agents-market-report

• Gartner – 40% of Enterprise Apps with AI Agents by 2026
  https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025

• Darktrace – State of AI Cybersecurity 2026 (92% Concerned)
  https://www.darktrace.com/blog/state-of-ai-cybersecurity-2026-92-of-security-professionals-concerned-about-the-impact-of-ai-agents

• Satya Nadella on “Managers of Infinite Minds” (Agents) – GeekWire
  https://www.geekwire.com/2026/satya-nadellas-new-metaphor-for-the-ai-age-we-are-becoming-managers-of-infinite-minds/

• Arvind Krishna / IBM on AI Operating Model – IBM Newsroom (Think 2026)
  https://newsroom.ibm.com/2026-05-05-think-2026-ibm-delivers-the-blueprint-for-the-ai-operating-model-as-the-ai-divide-widens