RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems

RecHarness is an agentic recommender optimization framework that separates direction selection from concrete code edits.

A bandit router uses past validation results to choose the next modification direction, while an LLM generates a hypothesis and code change within that direction.

The system adds a jump-basin mechanism to escape stalled local search and improve stability under limited experiment budgets.

It was tested across multiple recommendation tasks and backbones, including a 7-day online A/B test on a short-video advertising platform.

The approach aims to reduce manual engineering and outperform direct LLM-driven search in practical recommender tuning.