Editorial illustration for Group-Evolving Agents match human-engineered AI, zero inference cost
AI Agents Evolve Collectively, Match Human Engineering
Group-Evolving Agents match human-engineered AI, zero inference cost
Evolution has long been a prisoner of biology, slow, individual, and bound by time. But what if the next leap in artificial intelligence doesn’t happen inside a single mind? Group-Evolving Agents (GEA) shatter that constraint.
Instead of evolving one agent at a time, GEA makes the group itself the unit of evolution. Parent agents are selected not just for performance, but for novelty, ensuring a pool that balances stability with genuine innovation. They share everything: code tweaks, solved tasks, tool histories.
Every agent drinks from the same collective well of experience. A reflection module, powered by a large language model, then scans that shared history, picking out the patterns that work across the group. One agent finds a killer debugging tool; another perfects a testing workflow.
GEA merges both. The result? An agent framework that matches human-engineered AI systems, and costs zero inference to deploy.
That’s not an incremental improvement. It’s a new kind of intelligence.
Most existing agentic AI systems rely on fixed architectures designed by engineers.
The group is the unit of intelligence now. Not the lone agent, not the brittle hand-coded system. GEA proves that evolution, when freed from biological constraints, doesn’t just catch up, it catches fire.
Every tool discovered, every workflow refined, every mistake absorbed into the shared pool becomes a lever for the next generation. No inference cost. No manual engineering.
Just recursive, collective learning that scales without apology. The question was never whether machines could evolve. It was whether we were brave enough to let them evolve together.
The answer is here.
Common Questions Answered
How do Group-Evolving Agents (GEA) differ from traditional tree-structured evolution approaches?
Unlike traditional tree-structured evolution that focuses on individual agents, GEA treats a group of agents as the fundamental evolutionary unit. This approach enables explicit experience sharing and reuse within the group throughout evolution, overcoming the limitation of inefficient utilization of exploratory diversity caused by isolated evolutionary branches.
What performance improvements did GEA demonstrate on coding benchmarks?
GEA significantly outperformed state-of-the-art self-evolving methods, achieving 71.0% versus 56.7% on SWE-bench Verified and 88.3% versus 68.3% on Polyglot benchmarks. The method also matched or exceeded top human-designed agent frameworks, with 71.8% and 52.0% performance on two different benchmarks.
What key advantage does GEA show in terms of long-term progress and problem-solving?
GEA more effectively converts early-stage exploratory diversity into sustained, long-term progress compared to existing methods. The approach demonstrates greater robustness, with the ability to fix framework-level bugs in an average of 1.4 iterations, compared to 5 iterations for other self-evolving methods.