harness

An empirical arXiv study examines /retro, a skill that lets coding agents review completed sessions and propose harness changes, finding that while many proposals are generated, only some reduce repeated work on later tasks — showing harness self-evolution needs mechanisms to verify and revise changes.

DeepSeek announced Creator Mode in DeepSeek Harness, letting users turn ideas into custom plugins such as design workspaces, sandbox games, or 3D LLM visualizers.

The post explains the difference between AI agent frameworks (libraries like ADK, LangGraph, CrewAI) and agent harnesses (pre-built agents like Claude Code, Codex, Deep Agents), and offers guidance on when to choose each as the boundaries between them blur.

The author looked into Magpie, praising it as incredibly powerful — essentially the ultimate local model relay station that can route tokens pieced together from all over into any harness.

本文提出一种将评审者、过滤器与写作者分离的 harness 架构，通过上下文隔离减少 LLM 代理在扩写报告时的过度防御性写作，在 25 个商业项目草稿实验中将防御性命题从 59 降至 2（减少 96.61%），并将源主张保留率从 80.35% 提升至 93.12%。

kw-bench is a new benchmark built from real investigative knowledge-work engagements (e.g., five-year VAT reconstruction across SAP generations) that evaluates coding-agent harnesses with deterministic grading, finding that independent harnesses like OpenCode and micro-cc succeed where lab harnesses fail off their home model, and that frontier models such as Claude Opus 5.5 complete in minutes what weaker setups cannot.

A new product called Workbenches lets users lock in the best combination of harness, model, and tooling for specific classes of agent work, plugging into existing agent harnesses like Claude Code and shipping with benchmarks from day one.

作者对比了 Claude Code 与另外 11 个生产级 AI 编码代理的 harness 实现，指出三、四月的性能回退是 harness bug 而非模型问题，并总结出薄 harness 层的风险与改进建议。

Harness² 是 Google Research 推出的即插即用递归 agent 框架，让 LLM agent 在无人工验证的情况下自动优化其执行 harness（提示、工具、记忆），在 WorkBuddy-Bench 等专业基准上最高提升 11.5 个点。

Box CEO Aaron Levie 吐槽/点评一项关于 AI agent 架构的新研究，认为最终 agent 会自己确定其运行 harness，并引用 Pedro Domingos 称『模型吞掉了 harness』。

提出 ReHarness 框架，通过在严格在线流式协议下结合前向读写与后向修订（利用轨迹反馈修订检索到的反思），实现 LLM 智能体的持续学习，在三个基准和三种模型骨干上均优于基线方法。

A developer discusses the challenge of enabling communication between AI agents running on different harnesses and hosts, sharing experiences with file-based message passing and asking for recommendations on tools like Bolter that solve cross-harness agent messaging.

An in-depth breakdown of Lecture 5, "Agent Tools in Pictures," from Columbia University's Agentic Engineering course — 221 minutes covering the essence of tools, the complete call chain for a single invocation, five ways to package tools (functions / CLI / Skills / MCP / code execution), and security plus engineering judgment. The core thesis: "The model writes the order, the program does the cooking — prompts are not the lock, code is."

Voyager is an open harness for creative work, tuned so open models can integrate with creative tools like After Effects, DaVinci Resolve, and Figma, letting users collaborate with AI agents that learn their workflow via memory and custom skills.

Voyager 是面向创意工作的开放式 harness，被称作「创意领域的 Codex」，可驱动 Blender、DaVinci Resolve、Unity 等桌面应用，并支持 Opus、Astra、DeepSeek 等模型。评论指出通用编码 agent 的强大源于其 harness，创意工作同样需要专属 harness。

论文提出ReHarness，一种用于智能体持续学习的循环harness框架，在严格在线流式协议下通过backward revision补充前向信息流，从而在分布偏移下改进语言模型智能体的记忆与反思更新。

Meta 论文提出 RankEvolve，通过可执行操作协议（EOP）将研究阶段编译为受运行时强制执行的状态机，并让 Claude Code 与 Codex 作为节点互相审查修复，在预算持平下将执行准确率从 45.8% 提升到 62.5%，且在 MovieLens-20M 上改进推荐模型的 NDCG@10 达 4.48%。

Omar Sansevierino argues that the next wave of coding agents and harnesses hinges on extensibility and customization, citing Claude Code mods, DeepSeek Harness, and the newly shipped Pi 1.0 with Pi Durable as evidence that builders need more malleable, domain-specific harnesses.

Earendil and the Pi community shipped Pi 1.0 alongside an experimental new package, Pi Durable, a minimalistic framework for building long-running, malleable agents that can run anywhere with a JavaScript runtime and survive failures via pluggable storage backends.

LangChain shares a practical guide on building a model router inside the agent harness—covering task understanding, model selection, and outcome tracking—claiming a 64% reduction in median cost per thread with no quality loss.