Can an AI Filesystem unlock Intelligence? Agent Harness (Anthropic, Tsinghua)

Context Engineering is Not Enough: AI Agent Harnesses (Anthropic, Tsinghua Univ). LLMs suffer from "context rot." An LLM in a vacuum is just a stateless function. It only achieves long-horizon agentic intelligence when it can write its thoughts down, fork its context, and read its own history. The paper literally uses STATE_ROOT and manifest.json to prevent the AI from forgetting its own logic. The filesystem is the memory architecture. The authors (quote, see below): We introduce Natural-Language Agent Harnesses (NLAHs), which express harness behavior in editable natural language, and Intelligent Harness Runtime (IHR), a shared runtime that executes these harnesses through explicit contracts, durable artifacts, and lightweight adapters. All rights w/ authors: Natural-Language Agent Harnesses Linyue Pan1 Lexiao Zou2 Shuo Guo1 Jingchen Ni1 Hai-Tao Zheng1* from 1 Shenzhen International Graduate School, Tsinghua University 2 Harbin Institute of Technology (Shenzhen) @anthropic-ai @TsinghuaUniversity_official #aimemory #aiexplained #nextgentech #aiagents #filesystem #anthropic