RACL: Reasoning-Agent Control Layers for Continuous Metaheuristic Learning

Many organizations rely on metaheuristic optimizers for recurring operational tasks like vehicle routing, yet they often lack the expertise to adapt these tools as operational patterns evolve. The author introduces RACL (Reasoning-Agent Control Layer), a method that places a reasoning agent above an existing optimizer to manage its internal search behavior. Unlike traditional hyper-heuristics that might replace or reconfigure the solver, RACL acts as a supervisory layer that observes search history, formulates hypotheses, tests bounded interventions, and consolidates successful strategies into a reproducible policy.

The RACL cycle follows a structured process: observation, memory retrieval, reasoning, hypothesis generation, bounded experimentation, guardrail application, and policy consolidation. By treating each execution as evidence, the agent builds a memory of what works in specific contexts. Crucially, the agent is restricted from modifying business constraints (such as fleet capacity or delivery deadlines), ensuring that the optimizer remains responsible for feasibility while the agent focuses solely on improving search efficiency.

In experiments using a vehicle routing testbed, RACL demonstrated significant improvements over baseline strategies. It outperformed or tied a non-reasoning Stagnation-Triggered Policy (STP) in 18 out of 21 cases, with an average cost reduction of 0.641%. When compared to a fixed baseline, RACL achieved an average cost improvement of 8.337% in the measured sample. The study highlights that these gains were achieved without material computational overhead, and the agent was able to provide business-readable explanations for its control decisions, bridging the gap between technical optimization and operational transparency.