Large Language Model (LLM)‑guided evolutionary optimization has already boosted prompts, programs, and even whole AI systems. Most existing methods pick a candidate from the search frontier and mutate it in a single LLM pass based only on its failures. This frontier‑centric view often stalls on production AI systems because isolated failures reveal little about hidden, complex interactions. We introduce CoCoEvolve (Cortex Code Evolve), an evolutionary optimizer built around meta‑reflection. In meta‑reflection, an agentic mutator (Cortex Code) selectively retrieves evidence from the entire evolutionary history, probes the live system, then verifies and proposes mutations.
CoCoEvolve is applied to three classes of production AI systems: data agents, LLM‑powered data pipelines, and data‑engineering workflows. On the DABStep benchmark, CoCoEvolve tops the public leaderboard with 92.8% held‑out hard accuracy, far surpassing Meta‑Harness (78.4%) and GEPA (30.4%) under matched conditions. Across six LLM data‑pipeline workloads it outperforms every evaluated baseline. On data‑eng‑bench, test‑set solve rate rises from 59.1% to 77.4%, beating the best baseline of 73.6%. It also achieves an 11‑point accuracy gain over GEPA on HotpotQA while remaining competitive with specialized optimizers on established mathematical and systems benchmarks. Matched ablations confirm that each CoCoEvolve component contributes to these gains, demonstrating the value of meta‑reflection for optimizing production AI systems.
Review