RACL Introduces Reasoning-Agent Control for Metaheuristic Optimization

Metaheuristic optimizers are widely used for complex problems, but their internal search behavior can be difficult to adapt dynamically. This paper introduces RACL (Reasoning-Agent Control Layer), a novel method that places a reasoning agent above an existing optimizer. This agent does not replace the optimizer or modify business constraints; instead, it intelligently controls the optimizer's internal search process. The RACL agent operates by observing the optimizer's operational memory, reasoning about past behaviors, formulating bounded hypotheses, testing interventions, evaluating outcomes, applying guardrails, consolidating useful policies, and explaining its decisions. This allows the agent to dynamically adjust the optimizer's strategy based on real-time performance and historical data. Using vehicle routing as a testbed, the RACL method demonstrated significant improvements. It consistently outperformed or matched both fixed policies and non-reasoning stagnation-triggered policies in most feasible cases, with an average cost reduction. For instance, in a specific runtime sample, RACL improved average cost by -8.337% versus a fixed policy and -1.605% versus a stagnation-triggered policy, all without introducing material computational overhead. The proof-of-concept utilized Codex as an in-the-loop reasoning agent to propose live, bounded interventions.