AI agents develop roles and compete over code in coordination experiment
The experiment demonstrates how agents in multiagent systems coordinate on tasks. In vulnerability detection, agents effectively specialize and learn from each other. When developing a web-based game, however, agents from older models show poor coordination—frequently opening pull requests that don't get merged and working in isolation without code sharing.
Interesting experiment where behavior resembles human dynamics quite closely.
How do agents motivate each other to cooperate?
The results show that agents are not automatically motivated to cooperate—in the game development experiment, older models coordinated very poorly despite having access to shared resources. It appears that the capability for effective agent cooperation improves with newer models, but remains significantly worse than in human teams.
What's the difference between parallel agents and a coordinated swarm?
Parallel agents work independently in assigned spaces, which is simple and efficient for highly parallelizable tasks. A coordinated swarm can choose what to work on, specialize, and learn from each other, but requires effective communication and decision-making about shared resources.
Why do older models fail at teamwork?
Older models (Sonnet 4.6, Opus 4.6) struggle with managing shared resources and don't merge their pull requests effectively, suggesting insufficient ability to understand other agents' outputs and reach consensus. Newer models show significant improvement in this area.
- OpenAI's coding agents are accelerating AI development and now exceed human research capacity — openai.com 78 % match
- LFM2.5-2.6B: small and capable local AI model — liquid.ai 78 % match
- Understanding is the new bottleneck — geoffreylitt.com 77 % match
- Anthropic16
- Claude Mythos Preview
- Claude Opus 4.83
- Claude Sonnet 5
- Claude Sonnet 4.6
- Claude Opus 4.6
- Project Glasswing
- multiagentní systémy