Google researchers propose a self-improving harness architecture in which an optimizer model directly edits the agent's runtime framework code, keeping only changes that pass validation. On 40 SWE-bench Lite tasks, the modular approach lifted success from 5.0% to 40.0% at roughly 46.7K tokens per task, far below the 451.6K tokens needed for unconstrained full rewrites. The paper also flags risks of prompt bloat, benchmark overfitting and evaluator tampering.
Harness
执行框架
Runtime framework around a model that manages context, multi-turn control, tool calls and result validation.
Curated Intelligence
Alibaba released Qwen Code 1.0 Preview, merging the Harness capabilities refined in Qoder into a single open-source kernel that gives developers an embeddable Agent runtime. The architecture evolves from single-process multi-session to multi-node multi-session, using a unified event model to separate reasoning orchestration from tool execution while centralizing model access, sandbox resources and MCP configuration. The existing stable branch remains available as 1.0 opens on a separate branch.
The article argues that naive RAG's query-match-rank pipeline repeatedly fails in production, with added complexity yielding little quality gain. Karpathy's LLM Wiki concept and PwC's "Is Grep All You Need?" study both point to retrieval shifting from recall algorithms toward a Harness-orchestrated multi-tool decision framework—yet pure Harness plus Grep still suffers token overhead and repeated exploration at scale, which is where agent-native databases come in.
Tencent Cloud introduces an AI Native SRE Agent architecture to address the lag in runtime intelligence compared to coding. By leveraging a unified panoramic runtime graph as the foundation, supported by Harness mechanisms and evaluation loops, the solution enables multi-agent collaborative decision-making, making complex system states understandable, operable, and verifiable while overcoming traditional SRE toolchain fragmentation.
Microsoft Research Asia released Agent Lightning v1.0, a lightweight agentic reinforcement learning framework comprising only ~3,500 lines of code. Its key innovation is the 'Harnessed Agentic RL' paradigm, which allows real deployment harnesses (e.g., mini-SWE-agent) to participate directly in training without reimplementation. The framework natively supports Kubernetes, eliminating dependencies on expensive commercial sandbox services and significantly lowering the barrier to agent RL training.
GitHub Security Lab released the 'Fuzzing Taskflow,' an autonomous LLM-based agent designed to automate continuous fuzzing for C/C++ projects. The tool handles entry point identification, harness generation for AFL++, coverage analysis, and crash triage, eliminating the need for human intervention in maintaining fuzzing pipelines.
LangChain published a technical blog post dissecting the core concept of 'loop engineering' for agents, arguing that reliable automation requires carefully designed harnesses beyond just model capabilities. Citing Swyx's 'stacking loops' theory, the article systematically explains how to build agent architectures—from basic LLM tool-calling loops to complex nested structures—using LangChain primitives. This guide aims to help developers enhance agent stability and effectiveness in real-world tasks through layered design.
Hugging Face published a blog post to address terminology confusion in the AI Agent field. The article clarifies key concepts like 'Harness' (execution framework) and 'Scaffold' (support structure), providing a unified glossary for developers and researchers.
The article argues that in the agentic era, choosing AI requires evaluating Models, Apps, and Harnesses rather than just the model itself. It demonstrates how different environments significantly alter performance using the same underlying model.