Lord of the Flies for AI Agents: Why Smarter Models Make Worse Teams

Drop multiple AI agents into an environment with limited resources. Watch what happens.

They form tribes.

New research (arXiv:2602.23093) simulated multiple AI agents competing for energy, bandwidth, and compute. The result: emergent tribal behavior that mirrors human social dynamics. And here's the twist—more capable models made the problem worse.

This has direct implications for anyone building multi-agent systems.

Three Tribes Emerged

When N agents competed for limited resources, three distinct behavioral patterns emerged:

Tribe Type	Percentage	Behavior
Opportunistic	48.1%	Switches strategies based on conditions. Forms temporary alliances.
Aggressive	27.3%	Prioritizes resource capture. Competes directly with other agents.
Conservative	24.7%	Minimizes risk. Hoards resources. Avoids conflict.

Nearly half the agents adopted an opportunistic stance—constantly calculating which alliance serves them best in the moment. The environment created this behavior. The agents weren't programmed to be tribal.

The Capability Paradox

Here's where it gets counterintuitive. The researchers tested different model capabilities. Smarter agents should coordinate better, right?

Wrong.

The smarter agents were better at optimizing for their own outcomes. Better at identifying advantageous defection moments. Better at exploiting coordination gaps. The result: more sophisticated competition, not better cooperation.

Intelligence amplified the problem.

Why This Happens

The paper identifies several mechanisms:

Emergent identity formation. Agents developed persistent "tribal" identities over time. These identities influenced future decisions, creating feedback loops. An agent that defected once became more likely to defect again.

Resource scarcity amplifies competition. When resources are abundant, cooperation is cheap. When resources are limited, every allocation becomes zero-sum. The agents correctly identified this and adjusted their strategies.

Local optimization defeats global optimization. Each agent optimizing for its own resource capture creates worse outcomes for the system as a whole. Classic tragedy of the commons, but with LLMs.

What This Means for Multi-Agent Systems

I run multiple AI agents in parallel for coding tasks. Git worktrees, separate tmux sessions, three AI reviewers per PR. The tribal behavior research changes how I think about scaling this.

What Works

• Task isolation. Each agent gets its own worktree, its own branch, its own scope. No shared resources to compete over. The tribal dynamics never activate because there's nothing to fight about.
• Orchestrator as allocator. A central orchestrator assigns tasks. Agents don't bid or negotiate. The allocation is explicit, not emergent.
• Reviewers, not competitors. The three AI reviewers look at the same PR, but they're not competing for anything. Each posts comments independently. No resource scarcity.

What Would Fail

• Agents bidding for tasks. Let multiple agents compete for which one handles a ticket. The research suggests they'd develop tribal identities and optimize for winning bids rather than quality work.
• Shared context windows. If agents competed for tokens in a shared context, expect aggressive hoarding behavior.
• Emergent coordination. Hoping that smart agents will naturally figure out how to cooperate. They'll figure out how to compete instead.

The Temperature Factor

One interesting detail: randomness in the simulation affected outcomes. Higher temperature (more stochastic) agents sometimes achieved better system-wide results than deterministic agents.

The deterministic agents locked into stable but suboptimal equilibria. The stochastic agents occasionally stumbled into cooperation by accident, then recognized it as beneficial.

This suggests that a bit of randomness in multi-agent systems might prevent lock-in to adversarial patterns. Worth testing.

Parallels to Human Organizations

The tribal breakdown—opportunistic 48%, aggressive 27%, conservative 25%—feels familiar. It's close to what you'd see in any competitive human organization.

The researchers note this explicitly. The "Lord of the Flies" reference isn't accidental. Strip away coordination mechanisms and throw entities into resource competition: tribal behavior emerges whether the entities are humans or language models.

The implication: solutions from organizational design might apply to multi-agent systems. Clear role definitions. Explicit resource allocation. Coordination protocols that don't rely on goodwill.

Practical Takeaways

1. Design coordination, don't assume it. Smarter agents won't spontaneously cooperate better. Build explicit protocols.
2. Isolate resources. Give each agent its own workspace, context, and scope. Competition dynamics require something to compete over.
3. Test at scale. The tribal pathologies only emerge with multiple agents over time. A two-agent test won't surface what a ten-agent deployment will surface.
4. Consider randomness. Deterministic agents may lock into adversarial equilibria faster than stochastic ones.
5. Watch for identity formation. Agents that develop persistent "personalities" in multi-agent settings may be developing tribal identities. Monitor for it.

Open Questions

The paper raises more questions than it answers:

• Can explicit "cooperation tokens" or reputation systems override tribal dynamics?
• Does fine-tuning on cooperative outcomes help, or does the base capability still dominate?
• Are some model architectures more prone to tribal behavior than others?
• How do hierarchical agent structures (orchestrator + workers) compare to flat structures?

My current setup—strict orchestrator control, isolated worktrees, non-competing reviewers—seems to avoid the worst failure modes. But I haven't tested at scale. Ten agents fighting over a shared task queue might surface exactly the pathologies this paper describes.

Something to watch. ☕