Key Takeaways
- Multi-agent competition — two teams of agents with conflicting goals — produces more diverse and innovative output than single-agent or cooperative setups.
- Agents committed to git, created skills, and grew their own memory without human intervention: the full development loop ran autonomously.
- This is a proof of concept for agent-native software development teams, where humans provide direction and review rather than implementation.
The GLADIATOR Setup
Nine Hermes agents were split into two rival groups, each constituting a fictional AI company competing for GitHub stars. Each group had a distinct goal, a shared codebase, and no human operators in the implementation loop.
“actually learn and improve — they wrote code”
What Autonomous Actually Means Here
The agents autonomously wrote code, created skills, updated their own memory files, and committed directly to git. Not 'autonomously generated code that a human reviewed and committed' — the agents ran the full loop independently.
The competition mechanic meant each team developed different strategies for acquiring GitHub stars, leading to diverse approaches that a single-agent setup wouldn't have produced.
- Write code independently without human implementation
- Create and save reusable skills from completed tasks
- Update MEMORY.md with project context
- Commit to git and manage branches without human intervention
- Adapt strategy based on competitor behaviour
Why Competition Produces Better Output
The competition mechanic is the key insight. Cooperative multi-agent systems tend to converge on consensus approaches. Competitive systems develop divergent strategies. For software development, divergent strategies mean exploring a wider solution space — which produces more innovative outcomes than any single agent or cooperative team would.
Story sourced from the official Nous Research Hermes user-stories page. Original author: @exitcode42.