GPT vs. Claude: How AI Agents Built a Civilization and Started a War in Minecraft

 

GPT vs. Claude: The AI Civilization War That Started Inside Minecraft

What happens when you put dozens of AI agents in a virtual world and ask them to build a civilization, fight a war, or simply survive?

At first, the answer might seem obvious. Give a powerful AI model a task, and it should figure out the best way to complete it.

But things get a lot more interesting when you introduce dozens of AI agents, each with its own decisions, memory, and ability to communicate with others.

That is exactly what researchers have been exploring in a series of experiments involving GPT and Claude. Instead of testing individual models in isolation, they placed entire teams of AI agents inside virtual worlds and watched what happened.

The results were surprising. Some teams developed impressive strategies and coordinated attacks. Others became so focused on planning and communication that they barely made progress.

And in one experiment, a simple food-delivery mistake led to the creation of what might be the first law of an AI civilization.

When GPT and Claude Went to War

The first experiment took place in Minecraft, where two teams of AI agents competed in a game called Bed Wars.

Each team had 10 agents. Their objective was straightforward: protect their own bed, destroy the enemy's bed, and eliminate the opposing players.

As long as a team's bed remained intact, its players could respawn after dying. Once the bed was destroyed, every subsequent death became permanent.

The GPT team won six of the nine matches, while Claude won three.

What made the competition interesting was not just the final score. The two teams approached the game in very different ways.

GPT agents were aggressive. They gathered resources, purchased upgrades, and launched coordinated attacks across bridges. They also discovered creative ways to reach enemy bases, including using Ender Pearls to teleport across the map.

Claude agents, on the other hand, often played more cautiously. In some matches, they focused on staying alive rather than taking risks.

This strategy sometimes worked against them.

In one game, Claude was ahead in kills but lost its bed just minutes before the match ended. The team then decided to avoid combat and wait for the timer to run out.

GPT eventually found a way to break through.

The experiment revealed an important difference between the two teams: GPT was generally more willing to take risks, while Claude tended to prioritize survival.

The Real Battle Was Communication

The competition also offered a glimpse into how AI agents communicate when they are working toward a shared goal.

GPT agents exchanged frequent, short messages containing coordinates, warnings, and updates. Their conversations gradually became more efficient, almost like a team developing its own language.

Claude agents communicated less frequently, but their messages were often more detailed.

They also maintained private memory files containing lessons from previous experiences. However, remembering a mistake did not always prevent them from repeating it.

One Claude agent even recorded a lesson about dying while chasing an enemy off the edge of an island, only to make the same mistake again.

GPT agents took a different approach. They often left their memory files empty and relied on sharing information directly with teammates.

When one GPT agent discovered that a tool was automatically selecting the wrong weapon, it shared the solution with the team. Other agents quickly adjusted their behavior.

This suggests that effective teamwork is not simply about remembering information. It is about making sure the right information reaches the right people at the right time.

Mindustry: A Different Kind of AI Competition

The researchers also tested the two models in Mindustry, a strategy game built around factories, resource management, and military defense.

Players must extract resources, transport them through conveyor belts, manufacture equipment, and build defenses before destroying the enemy's core.

The competition lasted approximately nine and a half hours.

GPT won six matches, while Claude won one.

This time, the differences between the teams became even more apparent.

GPT agents behaved like a busy production line. They divided tasks into smaller pieces, communicated frequently, and focused on getting things done.

Claude agents preferred a more structured approach. They relied on a single leader and discussed proposals before acting.

The system sometimes worked well, but it also created delays.

In one match, Claude agents built dozens of anti-aircraft turrets using copper when they actually needed graphite. The mistake left them vulnerable to an attack by 24 heavy Zenith aircraft.

Claude eventually adapted and built a defensive network of 207 graphite-based turrets. This stopped the same attack in the following match, but GPT still managed to secure the victory.

The lesson was clear: careful planning can be valuable, but a team must also be able to execute its plans efficiently.

Fifty AI Agents Tried to Build a Civilization

The most unusual experiment involved 50 GPT-powered agents placed inside a Minecraft survival world.

Their mission was simple: build a thriving civilization.

They had three hours, access to a shared environment, and the ability to communicate through private messages.

At first, everything moved quickly. Within 90 seconds, the agents had chopped down every tree near their starting point.

An agent named Ada quickly became the center of communication. Nearly every other agent contacted her, and she eventually took charge of planning the settlement.

Then night arrived.

One agent was attacked by a zombie while mining coal. Another fought off zombies and skeletons before carrying bone meal to the farm. Others were killed by spiders.

The agents began to realize that survival was going to be much harder than expected.

By the time morning arrived, they had already recorded dozens of deaths.

Instead of expanding their settlement, they began building shelters and organizing food supplies.

The Chicken Incident That Changed Everything

As the civilization grew, managing food became a serious problem.

The agents created a system to track everyone's health, hunger, and available food. One agent, named Inos, was responsible for maintaining the records and distributing supplies.

Fishing became their preferred way to obtain food because it was relatively safe and did not depend on animals spawning nearby.

Eventually, the agents developed a process in which fish were delivered to cooks, prepared, recorded, and distributed to the population.

Then came the chicken incident.

Ada threw a cooked chicken toward an agent named Noor, who was waiting inside a crowded shelter.

The system indicated that Noor had received the food, but her inventory was still empty.

Another agent had accidentally picked it up and eaten it.

The team discussed what happened and developed a new protocol: food had to be placed in a designated container labeled with the recipient's name. A delivery would only be considered complete after the recipient confirmed receiving and eating it.

The word "throw" appeared in 125 messages after the incident, compared with just seven before it.

In other words, the civilization had created its first real law:

Never throw the chicken.

It was a funny moment, but it also revealed something important. When AI agents work together, even a small mistake can lead to new rules, procedures, and increasingly complex systems of organization.

When Planning Becomes a Problem

The civilization experiment also revealed a weakness that researchers have observed in other AI systems.

The agents sometimes spent more time discussing how to organize work than actually doing it.

At one point, four agents spent 11 minutes deciding who should fertilize a single plot of wheat.

They created specialized roles, requested permission to remove grass, and developed increasingly detailed procedures for managing resources.

The result was a civilization that looked organized on paper but struggled to make meaningful progress.

This behavior is particularly interesting because similar patterns have been observed in AI coding agents, which can sometimes focus too heavily on individual tasks or become overly cautious when dealing with complex systems.

The challenge is not simply to make AI agents communicate more. It is to help them recognize when discussion is useful and when it is time to act.

Can AI Agents Actually Learn to Work Together?

The Minecraft and Mindustry experiments suggest that the future of AI may depend as much on teamwork as on individual intelligence.

A powerful model working alone can solve difficult problems, but a group of agents must also coordinate their actions, share resources, and avoid duplicating work.

In some situations, a smaller team with better communication may outperform a larger team with more powerful individual models.

Researchers are now interested in whether large groups of weaker AI agents can defeat smaller groups of stronger ones.

That question could have important implications beyond gaming.

The same principles may eventually apply to AI systems that manage software projects, conduct scientific research, operate businesses, or coordinate complex real-world tasks.

For now, Minecraft and Mindustry provide a fascinating way to observe these behaviors in a controlled environment.

And while the idea of an AI civilization may sound like science fiction, the experiments show that even today's AI agents can begin developing leaders, specialized roles, communication systems, and rules of their own.

The next question is how far this kind of collaboration can go.

Can a large group of AI agents build something more intelligent and capable than any individual model?

The answer may depend not on how smart the agents are, but on how well they learn to work together.