Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

Topics: Data Mining, LLM Readability, LLMO / GEO, Query Fan Out, Retrieval Augmented Generation (RAG)

This paper introduces Harness-1, a 20-billion parameter AI search agent trained using reinforcement learning (RL) inside a specially designed “stateful harness.” The core argument is that traditional search agents are forced to handle both intelligent search decisions and routine bookkeeping tasks simultaneously, which makes learning inefficient. Harness-1 solves this by moving all mechanical record-keeping (such as tracking which documents have been seen, organizing evidence, and managing memory) out of the AI model and into the environment itself. Tested across eight different retrieval benchmarks covering web search, finance, patents, and multi-step question answering, Harness-1 achieves strong average recall performance, outperforming other open-source search agents of similar or larger size and remaining competitive with much larger commercial frontier models.