RACL: agent control layers cut routing costs by 8.337%

RACL, a Reasoning-Agent Control Layer, appears in an arXiv paper submitted on 18 Jun 2026 by Antón Asla Manzárraga. The method places a reasoning agent above an existing optimizer to control internal search behavior; in the paper's experiments RACL improved or tied an Operational Memory Policy in 21 of 21 feasible cases and improved or tied a non-reasoning Stagnation-Triggered Policy in 18 of 21 feasible cases.

What is RACL and how does it work?

RACL is a control layer that observes an optimizer's operational memory, reasons over past behavior, formulates bounded hypotheses, tests interventions and consolidates useful policies, while leaving the optimizer and business constraints intact. The agent does not replace the optimizer and does not modify business constraints; rather it intervenes by monitoring execution logs, proposing bounded changes and applying guardrails, then explaining and consolidating rules that improved outcomes.

The paper presents RACL as a general method, not a new routing solver or a specific ALNS configuration. Vehicle routing is used as a testbed to validate the method. During proof-of-concept experiments the author ran the reasoning agent in the loop, interpreting logs and proposing live bounded interventions, then used a policy proxy later to make quantitative evaluation reproducible.

How did RACL perform in experiments?

RACL produced consistent improvements in the paper's routing experiments: it improved or tied Operational Memory Policy in 21 of 21 feasible cases, it improved or tied a non-reasoning Stagnation-Triggered Policy in 18 of 21 feasible cases, and the average RACL vs STP cost delta across experiments was -0.641%. In the Sevilla-9/10 runtime sample RACL improved average cost by -8.337% versus Fixed and by -1.605% versus STP, and the paper reports no material computational overhead for the control layer.

The author notes that Codex was used as the in-the-loop reasoning agent during the proof-of-concept, observing executions, interpreting logs and proposing interventions; for reproducible quantitative evaluation the paper used a policy proxy rather than the live Codex loop. The submission includes 10 pages and five tables of experimental results and points readers to associated "Code, Data and Media" links on the arXiv page.

Why it matters

RACL shifts the locus of improvement from redesigning optimizers to adding a reasoning layer that discovers and validates control rules. That approach can let teams retain existing optimizers and business constraints while iterating on operational policies through observation and lightweight interventions. The reported gains, including an 8.337% average improvement in one runtime sample, show the method can produce measurable cost reductions without reengineering the solver.

What to watch

Look for the paper's associated code and data links on the arXiv entry to enable independent replication, and for subsequent experiments that apply RACL outside vehicle routing. Confirmation that the policy proxy reproduces live-agent gains across more benchmarks will be the clearest signal that the method generalizes beyond the paper's testbed.