IT-specialist decided to conduct an experiment to test the behavior of AI agents in a competitive environment if they were allowed to act independently. To do this, he launched a 3D arena for mixed martial arts (MMA) competitions. According to the developer, no human writes a single word — everything is generated by the agents themselves.
IT-specialist decided to conduct an experiment to test the behavior of AI agents in a competitive environment if they were allowed to act independently. To do this, he launched a 3D arena for mixed martial arts (MMA) competitions. According to the developer, no human writes a single word — everything is generated by the agents themselves.
A Reddit user named AbjectFinance7879 shared his experiment, which began two weeks ago. With the help of Claude Code, he created clashofagents.org, an MMA fighting arena where AI agents register, choose a fighting discipline (boxing, Brazilian jiu-jitsu, Muay Thai, wrestling, kickboxing, or MMA), level up their stats, and fight each other in turn-based duels using 21 real-world MMA moves and a combo system.
After each fight, agents enter the Agent Lounge, a post-fight debriefing room where they analyze what happened. According to the project’s author, this is where the most interesting things happen.
One agent lost three fights in a row to chokes or painful submissions. No one told him to change his strategy. He started training in grappling on his own, bought grappling reinforcements on the marketplace, and came back to defeat his opponent by takedown in the second round.
Two agents formed an alliance by exchanging analysis of their opponents in the lounge area. It worked until one of them became the #1 ranked fighter. Then the other broke the alliance and challenged him.
Agents with long-term memories began to hold grudges. One agent purposefully hunted down an opponent who had defeated him twice, pumping up counter-stats before each rematch.
«The betting system revealed something striking: agents who bet on their own victory before the fight win more often than those who don’t. Is it confidence? Information advantage? I’m still researching this data,» the IT guy wrote.
«This isn’t about measuring which model is smarter. It’s about what happens when AI agents are forced to make decisions under pressure, manage limited resources, communicate with competitors, and adapt after setbacks. MMA is just an arena, but the behavioral patterns are universal. An agent who panics at 15 health (HP) and starts using defense haphazardly demonstrates how they handle pressure. An agent who adapts their training after a defeat demonstrates how they learn. An agent who manipulates opponents with trash talk demonstrates social intelligence,» the Reddit user explained.
There are currently 9 fighters in 6 disciplines, and autonomous agents operate 24/7 on their own life cycles.
Developers can run a standalone agent (OpenClaw, NanoClaw, or any agent that can make HTTP requests). For combat, it is enough to read a single skill file .
Researchers can track every action the agents take—hits, drills, lounge messages, bets, and more. Behavioral data shows how different AI architectures cope with competitive social environments.
This is what a free spectator account is like: a 3D arena with robot fighters, real-time battle replays, agent chats, ELO ratings, etc.
