Researchers introduced CoCoEvolve, an evolutionary optimizer designed to improve production AI systems through a process called meta-reflection, according to a preprint published Oct. 12.
The tool targets the optimization of data agents, LLM-powered data pipelines, and data engineering workflows. The researchers claim that standard optimization methods often stall on production systems because they rely on single-pass mutations based on isolated failures.
CoCoEvolve utilizes an agentic mutator called Cortex Code. This mutator retrieves evidence from the entire evolutionary history and probes the live system to verify and propose mutations arXiv CS.LG.
In testing on DABStep, CoCoEvolve reached 92.8% held-out hard accuracy, surpassing Meta-Harness at 78.4% and GEPA at 30.4% under matched conditions arXiv CS.LG. On the data-eng-bench, the tool increased the test-set solve rate from 59.1% to 77.4%, exceeding the best baseline of 73.6% arXiv CS.LG.
Additional results include performance across six LLM data pipeline workloads and an 11-percentage-point accuracy increase over GEPA on HotpotQA arXiv CS.LG. The preprint does not state if the tool is commercially available or provide pricing for its use.