A research team has built a self-revising ensemble of AI coding agents that competes internally to improve itself — no human tuning required.
The framework, called Evolutionary Ensemble (EvE), organizes existing coding agents into two co-evolving populations: one focused on solving code problems, the other on refining the guidance those agents follow. Agents race against each other in real time, earning or losing Elo ratings based on how much they improve the current best solution. The system then updates the guidance states that shape future agent behavior, creating a feedback loop that tightens with each cycle. Researchers tested it on a genuine research bottleneck in a method called In-Context Operator Networks, where EvE independently found a rescale-then-interpolate technique that let models handle varying numbers of examples reliably.
What separates EvE from the crowded field of "LLMs as optimizers" work is that it leaves the underlying models alone and evolves the instructions instead. That matters because swapping in a better base model is expensive and slow; rewriting guidance is cheap and continuous. The ablation tests are the most interesting part: a frozen "best-evolved" agent still plateaued, while the live-evolving ensemble kept climbing — suggesting that adaptation, not capability, is the actual ceiling.
The result is a framework that treats agent coordination as an evolutionary problem rather than an engineering one — a bet that self-organization will outrun hand-tuned pipelines, though whether that holds outside controlled benchmarks remains to be seen.