EvE framework lets coding agents co-evolve to discover algorithms

Evolutionary Ensemble (EvE) is a decentralized framework that moves beyond the typical 'LLMs as optimizers' paradigm. Instead of tweaking the base agent's LLM, EvE fixes the agent substrate and focuses on evolving the cumulative guidance and skills that dictate agent behaviors. The system maintains two co-evolving populations: functional code solvers (which generate algorithmic solutions) and agent guidance states (which modify how agents approach problems). These populations are evaluated through a synchronous race, with each agent's empirical Elo rating updated based on the marginal gains it contributes to the current solver state. This creates a live, self-revising ecosystem where agents continuously adapt to the shifting search landscape.

EvE was applied to a research bottleneck in In-Context Operator Networks (ICON), where it autonomously discovered a robust rescale-then-interpolate mechanism that enables reliable example-count generalization. Crucially, controlled ablations revealed that stage-dependent agent adaptation is absolutely necessary—variants driven by a fixed initial agent or even a frozen 'best-evolved' agent suffered from phase mismatch and static performance ceilings. Only the full EvE ensemble, with dynamic guidance updates, broke through these barriers. The paper demonstrates that organizing existing coding agents into a co-evolving ensemble is the fundamental driver for algorithmic discovery, offering a scalable path for AI to autonomously improve complex codebases.