# Harness Engineering for Self-Improvement # 用于自我提升的Harness（运行框架）工程 **Date:** July 4, 2026 | **Estimated Reading Time:** 28 min | **Author:** Lilian Weng --- The concept of **recursive self-improvement (RSI)** dates back to [I. J. Good (1965)](https://philpapers.org/rec/GOOSCT), a British mathematician who worked alongside Alan Turing. In his foundational work, he defined an “ultraintelligent machine” as a system that can surpass humans in all intellectual activities and design better machines to improve itself, potentially leading to an "intelligence explosion" (often referred to as the singularity). Later, prominent AI safety researcher [Yudkowsky (2008)](https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement) used the phrase “recursive self-improvement” for a specific feedback loop: an AI uses its current intelligence to improve the cognitive machinery that produces its intelligence. **递归自我提升（Recursive Self-Improvement, RSI）** 的概念可以追溯到 [I. J. Good (1965)](https://philpapers.org/rec/GOOSCT)。作为曾与阿兰·图灵共事的英国数学家，他在其奠基性著作中将“超智能机器（ultraintelligent machine）”定义为一种能在所有智力活动中超越人类，并能设计出更好的机器来提升自身的系统，这可能最终导致“智能爆炸”（即通常所说的技术奇点）。后来，著名的AI安全研究员 [Yudkowsky (2008)](https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement) 使用“递归自我提升”一词来描述一个特定的反馈循环：人工智能利用其当前的智能来改进产生其智能的认知机制。 This feedback loop in modern AI may indicate the model rewriting its own weights directly, or more broadly the model improves the *training pipeline* and the *deployment system*, which in turn enables a better successor model with improved performance across economically valuable tasks. The speed of research development in AI has been shown to drastically accelerated in frontier labs like [Anthropic](https://www.anthropic.com/institute/recursive-self-improvement) and [OpenAI](https://openai.com/index/how-agents-are-transforming-work/), where AI agents are already assisting researchers in writing code and running experiments. 在现代人工智能领域，这种反馈循环可能意味着模型直接重写自身的权重，或者更广泛地说，模型通过改进**训练流水线（training pipeline）** 和 **部署系统（deployment system）**，进而催生出一个在各种具有经济价值的任务上表现更优的后继模型。在诸如 [Anthropic](https://www.anthropic.com/institute/recursive-self-improvement) 和 [OpenAI](https://openai.com/index/how-agents-are-transforming-work/) 这样的前沿AI实验室中，人工智能的发展速度已经显著加快，AI智能体（Agents）已经开始协助研究人员编写代码和运行实验。 I explicitly mention *“deployment system”* because the layer between the raw model and the real-world context seems to be as important as the model’s raw intelligence (i.e. the evals right after pretraining). Harnesses are important components of AI deployment, as shown by successful coding agent products such as Claude Code and Codex. In traditional software engineering, a test harness is a collection of software and test data configured to test a program unit. In the context of AI, a **harness** (or scaffolding) is the overarching system surrounding a base model that orchestrates execution and decides how the model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results. 我在这里特别强调 **“部署系统”**，是因为介于底层基础模型与真实世界上下文之间的这一中间层，似乎与模型本身的原始智能（即预训练后立刻进行的评估表现）同等重要。正如 Claude Code 和 Codex 等成功的编程智能体产品所展示的那样，Harness（运行框架/脚手架）是AI部署的重要组成部分。在传统软件工程中，Harness 通常指用于测试程序单元的软件和测试数据集合；而在人工智能语境下，**Harness（运行框架）** 是指围绕基础模型构建的整个外围系统，它负责编排执行流程，并决定模型如何思考与规划、调用工具与执行动作、感知与管理上下文、存储产出物（artifacts）以及评估结果。 This one post will focus on research around harness engineering and how it contributes to RSI. Much recent work on auto-research, self-improving agents, and evolutionary program search can be organized around this question. Other work on model self-play, synthetic data, test-time training and a broader theme of continual learning also matches the RSI vision (e.g. [Yuan et al. 2024](https://arxiv.org/abs/2401.10020), [Chen et al. 2024](https://arxiv.org/abs/2401.01335)), [Zhao et al. 2025](https://arxiv.org/abs/2505.03335), [Choi et al. 2026](https://openreview.net/forum?id=lTbBFAoPSA)) but they will not be the focus of this post. 本文将重点探讨围绕 Harness 工程的研究及其对 RSI 的贡献。近期许多关于自动化研究（auto-research）、自我提升智能体（self-improving agents）以及演化程序搜索（evolutionary program search）的工作都可以围绕这个问题来梳理。其他关于模型自我博弈（self-play）、合成数据（synthetic data）、测试时训练（test-time training）以及更广泛的持续学习（continual learning）主题的工作也符合 RSI 的愿景（例如 [Yuan 等 2024](https://arxiv.org/abs/2401.10020), [Chen 等 2024](https://arxiv.org/abs/2401.01335)), [Zhao 等 2025](https://arxiv.org/abs/2505.03335), [Choi 等 2026](https://openreview.net/forum?id=lTbBFAoPSA)），但这些并非本文的重点。 ## Harness Design Patterns ## Harness（运行框架）设计模式 Compared with [early agent frameworks](https://lilianweng.github.io/posts/2023-06-23-agent/) like LangChain or AutoGPT where the paradigm was simply “agent = LLM + memory + tools + planning + action”, harnesses engineering additionally include *workflow design (e.g. loop engineering), evaluation, permission controls, and persistent state management*. It is no longer only prompt templates, but closer to runtime and software system design: how the model observes, acts, memorizes, checks itself, and improves. 与 LangChain 或 AutoGPT 等[早期智能体框架](https://lilianweng.github.io/posts/2023-06-23-agent/)（其范式简单地定义为“智能体 = LLM + 记忆 + 工具 + 规划 + 动作”）相比，Harness 工程还额外包含了**工作流设计（例如循环工程）、评估机制、权限控制以及持久化状态管理**。它不再仅仅是提示词模板（prompt templates）的拼接，而是更接近于运行时（runtime）和软件系统设计：即模型如何观察、行动、记忆、自我检查以及自我改进。 The design should be deliberately simple and generic to enable generalization, likely with reference to existing software engineering practices to benefit from prertaining knowlege. There is also a strong analogy between operating systems and harnesses. Similar to an OS, a harness should encapsulate complicated logic while keeping the interface simple. Meanwhile, configs, tool interfaces and other protocols may gradually become standardized across the industry. 该设计应该刻意保持简单和通用以实现泛化，很可能需要借鉴现有的软件工程实践，以便充分利用模型在预训练阶段学到的代码知识。操作系统（OS）与 Harness 之间存在着强烈的类比关系。与操作系统类似，Harness 应该封装复杂的底层逻辑，同时保持接口的简洁。与此同时，配置、工具接口以及其他协议可能会逐渐在整个行业中标准化。 ### Pattern 1: Workflow Automation ### 模式 1：工作流自动化 Defining a workflow in which the model can operate, test, and iterate is a key design for automation. Karpathy’s autoresearch repo (https://github.com/karpathy/autoresearch) is a clean example of how such a workflow can be constructed. A common workflow follows a goal-oriented loop of plan, execute, observe/test, improve, and execute again *until* the goal is achieved. The process may trigger proactive requests to users for clarity in task specification or execution preference. 定义一个允许模型在其中进行操作、测试和迭代的工作流，是实现自动化的关键设计。Andrej Karpathy 的 autoresearch 仓库（https://github.com/karpathy/autoresearch）提供了一个如何构建此类工作流的简洁示例。一个常见的工作流遵循以目标为导向的循环：计划、执行、观察/测试、改进，然后再次执行，**直到**目标达成。这个过程可能会主动向用户发起请求，以明确任务规范或执行偏好。

A simplified Codex agent loop: the agent calls tools and tool responses affect the model's next generation.
(Image source: OpenAI codex agent post)

简化的 Codex 智能体循环：智能体调用工具，工具的响应将影响模型的下一次生成。

The workflow graph also emphasizes the model analyzing its own trajectories and failure cases and then iterating on its progress through an “agent runtime” rather than a static prompt template. 上面的工作流图也强调了模型会分析自身的执行轨迹和失败案例，然后通过“智能体运行时（agent runtime）”不断迭代推进，而不是依赖静态的提示词模板。 ### Pattern 2: File System as Persistent Memory ### 模式 2：将文件系统作为持久化记忆 A recurring pattern in long-horizon agent systems is simple control over rich states and artifacts. A harness should not carry the entire workflow and all logs in context; instead, it should keep durable state in files. In long-horizon agentic rollout, artifacts such as experiment logs, code diffs, paper summaries, error traces, and past rollout trajectories often grow much longer than the context window that the model has trained for. 在长周期（long-horizon）智能体系统中，一个反复出现的模式是对丰富的状态和产出物进行简单的控制。Harness 不应将整个工作流和所有日志都塞进模型的上下文窗口中；相反，它应将持久状态保存在文件中。在长周期的智能体执行过程（rollout）中，诸如实验日志、代码差异（diffs）、论文摘要、错误追踪以及过去的执行轨迹等产出物，其长度往往会远远超出模型训练时所能容纳的上下文窗口限制。 Learning how to read, write, and edit the file system (commonly via `bash` commands) is a foundation skill for LLMs, and thus managing persistent memory in the simple form of files naturally benefits from improvements in core model capability. 学习如何读取、写入和编辑文件系统（通常通过 `bash` 命令）是 LLM 的一项基础技能，因此，以简单的文件形式管理持久化记忆，自然能从核心模型能力的提升中获益。 ### Pattern 3: Sub-agent and Backend Jobs ### 模式 3：子智能体与后台任务 A harness can spawn multiple subagents to execute in parallel and monitor backend jobs. This is useful when the main agent needs to search multiple hypotheses, run experiments concurrently, or delegate isolated subtasks without polluting the main context. The parent agent then needs a small process manager: launch jobs, inspect logs, cancel failed runs, and merge results back into the main agent thread. Harness 可以生成多个子智能体（subagents）并行执行，并监控后台任务。当主智能体需要搜索多个假设、同时运行多个实验，或委派孤立的子任务而不污染主上下文时，这一功能非常有用。此时，父智能体需要一个轻量级的进程管理器来：启动任务、检查日志、取消失败的运行，并将结果合并回主智能体线程。 The key design choice is to make parallelism explicit and inspectable. If subagent outputs only live in a transient chat context, they quickly become obselete and hidden. If they are stored as files, logs, and status records, the model can recover after interruptions and reason over its own execution history. 这里的关键设计选择是让并行化变得明确且可检查。如果子智能体的输出仅存在于短暂的聊天上下文中，它们很快就会变得过时并被隐藏。如果它们被存储为文件、日志和状态记录，模型就可以在中断后恢复，并基于自身的执行历史进行推理。 ### Case study: Coding Agent Harness ### 案例研究：编程智能体 Harness The core interface of mainstream coding agents has become stabilized across Claude Code, Codex, OpenCode, and Cursor-style agents. They commonly use a loop like: 主流编程智能体（如 Claude Code、Codex、OpenCode 以及类似 Cursor 的智能体）的核心接口已经趋于稳定。它们通常使用如下循环：

With access to a set of tools, the coding agent is able to develop and debug issues in a given repository, similar to how human developers are equipped with IDEs. 通过使用一系列工具，编程智能体能够在给定的代码仓库中开发和调试问题，这与人类开发者配备 IDE（集成开发环境）的方式非常相似。 *(Not a comprenhensive list; shown for demonstration. Read [this](https://github.com/yasasbanukaofficial/claude-code) if interested.)* *（以下并非详尽列表，仅供演示。如果感兴趣，可阅读[此文](https://github.com/yasasbanukaofficial/claude-code)。注：MCP指的是Model Context Protocol，由Anthropic开发的一种标准化工具调用协议）* | Group (工具组别) | Tool definitions (工具定义) | | :--- | :--- | | File system (文件系统) | - File discovery (文件查找): `glob`, `grep`, `ls`
- File read (文件读取): `read`, `read_many`
- File modification (文件修改): `write` (a whole new file/写入新文件); `edit` (string exact-match replacement/字符串精准替换); `multi_edit`; `apply_patch` (applies a structured patch/diff / 应用结构化补丁) | | Shell execution (Shell 执行) | Run commands (运行命令): `bash`, `PowerShell` | | IO (输入输出) | `lsp`, git tools like (git工具如) `git_status`, `git_diff`, `git_commit` | | External context (外部上下文) | MCP tools (模型上下文协议工具), Skills (技能) | | Web search (网络搜索) | `web_search`, `web_fetch`, browser tools (浏览器工具) | | Artifacts (产出物) | Read docs, images (阅读文档、图像); generate HTML, images (生成HTML、图像) | | Backend processes (后台进程) | Such as (例如): `CronCreate`, `CronDelete`, `CronList` | | Agent delegation (智能体委派) | Such as (例如): `spawn_agent`, `resume_agent`, `wait_agent`, `list_agents`, `close_agent`, `interrupt_agent`, etc. | ### Harness Layer vs Core Intelligence? ### Harness 层与核心智能的关系？ It is hard to forecast how much the future of RSI will rely on harness engineering, but the near-term path of RSI is unlikely to start as a model directly rewriting its weights. My prediction of a practical near-term path is: 很难预测 RSI 的未来在多大程度上将依赖于 Harness 工程，但 RSI 近期的发展路径不太可能直接从模型重写自身权重开始。我对近期切实可行路径的预测是： 1. Harness engineering will evolve in the direction of meta-methodology (i.e. improving the machinery for getting better answers, not just improving the answer itself). The harness system itself becomes an optimization target, with fewer heuristic rules and more general mechanisms. 2. In turn, mature harnesses enable auto-research for model self-improvement loop and smarter models prevents harnesses from overengineering and keep the system sustainable. 1. Harness 工程将朝着元方法论（meta-methodology）的方向演进（即改进获取更好答案的“机器/机制”，而不仅仅是改进答案本身）。Harness 系统本身将成为一个优化目标，启发式（手工设定）规则会减少，取而代之的是更通用的机制。 2. 反过来，成熟的 Harness 将推动模型自我提升循环的自动化研究（auto-research），而更智能的模型也能防止 Harness 被过度设计（overengineering），从而保持系统的可持续性。 Eventually it is possible that many harness improvements will be *internalized* into core model behavior, but the interface with external context and tools should remain. We have seen a softer version of this pattern with [prompt engineering](https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/): manual prompt tricks became less central as instruction tuning and model reasoning improved, but *the need to specify goals, constraints, context, and evaluation did not disappear*. 最终，许多 Harness 层的改进有可能会被**内化**为核心模型的行为能力，但与外部上下文和工具交互的接口仍将保留。我们已经在[提示词工程（prompt engineering）](https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/)的发展中看到了这种模式的温和版本：随着指令微调（instruction tuning）和模型推理能力的提升，手工堆砌提示词技巧变得不再那么核心，但**明确指定目标、约束条件、上下文和评估标准的需求并未消失**。 ## Harness Optimization ## Harness（运行框架）优化 The progression in the object being optimized in the harness system is roughly: instruction [prompts](https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/) → structured context → workflow → harness code → optimizer code. As the model becomes more intelligent and powerful, we move toward more complex targets and generic methods. 在 Harness 系统中，被优化的对象演进过程大致为：指令[提示词](https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/) → 结构化上下文 → 工作流 → Harness 代码 → 优化器代码。随着模型变得更加智能和强大，我们正朝着更复杂的目标和更通用的方法迈进。 ### Context Engineering ### 上下文工程 Simply appending all the tool responses and model generations into the context can quickly grow out of control as the agentic job horizon increases significantly. Context management is a layer to construct a more structed and concise context for LLM and manage persistant states. There is no doubt that long-context research will keep on making progress (with models now handling millions of tokens) but at the moment long-context intelligence and context engineering sometime intertwines to prevent issues like the "lost in the middle" phenomenon. 随着智能体任务周期的显著增加，如果简单地将所有工具响应和模型生成内容追加到上下文中，上下文长度很快就会失控。上下文管理是一层用于为 LLM 构建更结构化、更简洁的上下文并管理持久化状态的机制。毫无疑问，长上下文研究将继续取得进展（目前的模型已能处理数百万个 token），但在现阶段，长上下文智能与上下文工程有时会交织在一起，以防止诸如“迷失在中间（lost in the middle）”等现象的发生。 **Agentic Context Engineering** (ACE; [Zhang et al. 2025](https://arxiv.org/abs/2510.04618)) treats context as an evolving playbook rather than an increasingly lengthening prompt. It has three components to maintain one context playbook of bullet points, each with an identifier and a description. **智能体化上下文工程（Agentic Context Engineering, ACE**; [Zhang 等 2025](https://arxiv.org/abs/2510.04618)）将上下文视为一本不断演进的指导手册（playbook），而不是越来越长的提示词。它包含三个组件来维护这本由要点（bullet points）组成的上下文手册，每个要点都有一个标识符和一段描述。 1. *Generator*: produces task trajectories, with reference to bullet points. 2. *Reflector*: distills insights from successful and failed trajectories. 3. *Curator*: updates the structured context with incremental, itemized entries. 1. *生成器（Generator）*：参考各个要点，生成任务执行轨迹。 2. *反思器（Reflector）*：从成功和失败的轨迹中提取洞见（insights）。 3. *整理器（Curator）*：用增量式的、逐项列出的条目来更新结构化上下文。

The framework of Agentic Context Engineering (ACE). (Image source: Zhang et al. 2025)

智能体化上下文工程（ACE）框架。

To prevent context collapse and brevity bias during iterative rewrites, one key design choice in ACE is that the curator does not rewrite a full prompt blob. It instead outputs a collection of structured, itemized bullets in the form of (identifier, description), and these bullets are merged into a structured context logbook with deterministic logic. The context items are refined and deduplicated periodically. 为了防止在迭代重写期间发生上下文崩溃和“过度简短偏见（brevity bias）”，ACE 的一个关键设计选择是，整理器（curator）不会重写整个提示词块。相反，它输出的是一组以（标识符，描述）形式存在的、结构化的条目集合，然后通过确定性逻辑将这些条目合并到结构化的上下文日志簿中。这些上下文条目会定期进行优化和去重。 The fact that ACE learns insights from rollouts helps us move toward self-managed memory, but the update rules and the overall workflow are still handcrafted. To move toward a more self-improving loop, **Meta Context Engineering** (MCE; [Ye et al. 2026](https://arxiv.org/abs/2601.21557)) separates the mechanism (how to manage context) from the artifact content (what is in context), running skill evolution at the meta-optimization level and context optimization at the base level. ACE 能够从执行轨迹中学习洞见，这一事实帮助我们朝着自我管理记忆的方向迈进，但其更新规则和整体工作流仍然是手工编写的。为了走向更具自我提升能力的循环，**元上下文工程（Meta Context Engineering, MCE**; [Ye 等 2026](https://arxiv.org/abs/2601.21557)）将机制（如何管理上下文）与产出物内容（上下文里有什么）分离开来，在元优化层（meta-optimization level）运行技能演化，在基础层（base level）运行上下文优化。 An MCE skill $s \in \mathcal{S}$ defines a context function $c_s = (\rho_s, F_s)$ and maps an input $x$ to context $c = F_s(x; \rho_s)$, where: - $\rho_s = \{\rho_1, \dots, \rho_m\}$ are static components (prompts, knowledge bases, code libraries). - $F_s = \{F_1, \dots, F_k\}$ are dynamic operators (search, selection, filtering, formatting). MCE 的一个技能 $s \in \mathcal{S}$ 定义了一个上下文函数 $c_s = (\rho_s, F_s)$，并将输入 $x$ 映射到上下文 $c = F_s(x; \rho_s)$，其中： - $\rho_s = \{\rho_1, \dots, \rho_m\}$ 是静态组件（提示词、知识库、代码库）。 - $F_s = \{F_1, \dots, F_k\}$ 是动态算子（搜索、选择、过滤、格式化）。 The bi-level optimization is to find the best context $c_s^*$ given skill $s$ on the training data, while the outer loop finds the optimal skill that provides the best performance on the validation set: 这种双层优化的目标是：在内层循环中，给定技能 $s$，在训练数据上找到最佳的上下文 $c_s^*$；而在外层循环中，寻找在验证集上提供最佳性能的最优技能 $s^*$： $$ \text{Inner: } c_s^* = \arg\max_{c_s} J_{\text{train}}(c_s; s) \quad \text{Outer: } s^* = \arg\max_{s \in \mathcal{S}} J_{\text{val}}(c_s^*) $$ The skill database tracks the history of previous skills, context functions and eval metrics $\mathcal{H}_{k-1} = \{(s_i, c_i, J_i^{\text{train}}, J_i^{\text{val}})\}_{i=1}^{k-1}$. A meta-level agent performs agentic [crossover](https://en.wikipedia.org/wiki/Crossover_(evolutionary_algorithm)) over prior skills to create a new skill given a task $\tau$: $s_k = \text{crossover}(\tau, \mathcal{H}_{k-1})$. 技能数据库跟踪记录了以往技能、上下文函数和评估指标的历史：$\mathcal{H}_{k-1} = \{(s_i, c_i, J_i^{\text{train}}, J_i^{\text{val}})\}_{i=1}^{k-1}$。给定任务 $\tau$，元级别智能体会对先前的技能执行智能体化的[交叉操作（crossover，遗传算法中的概念）](https://en.wikipedia.org/wiki/Crossover_(evolutionary_algorithm)) 以创建新技能：$s_k = \text{crossover}(\tau, \mathcal{H}_{k-1})$。 Then a base-level context engineer executes the skill $s_k$ and learns the context function from rollout feedback $\mathcal{R}_k$, guided by the current skill: $c_k = \text{engineer}(\tau, s_k; c_{k-1}^*, \mathcal{R}_k)$. 接着，基础级别的上下文工程师会执行技能 $s_k$，并在当前技能的指导下，从轨迹反馈 $\mathcal{R}_k$ 中学习上下文函数：$c_k = \text{engineer}(\tau, s_k; c_{k-1}^*, \mathcal{R}_k)$。

The framework of Meta Context Engineering (MCE): meta-level skill evolution searches over context-management mechanisms, while the base level optimizes the task context. (Image source: Ye et al. 2026)

元上下文工程（MCE）框架：元级别的技能演化在上下文管理机制上进行搜索，而基础级别则优化任务上下文。

MCE does not enforce a heuristic rule for how to structure context as ACE does. It uses *free-form skills* to store the most important knowledge for a task, and evolves the skill and the skill-conditioned context iteratively together. Implementation-wise, a context function $c$ is instantiated as a collection of files in a dedicated directory, including both static (`skill.md`) and dynamic (context and data rollouts) components. Both meta-level and base-level optimization are executed in agentic coding envs with a standard tool set: MCE 不像 ACE 那样强制规定一种启发式规则来构建上下文。它使用**自由形式的技能（free-form skills）**来存储任务最重要的知识，并将技能与基于技能的上下文共同进行迭代演化。在实现上，上下文函数 $c$ 被实例化为专用目录下的文件集合，既包含静态组件（`skill.md`），也包含动态组件（上下文和数据轨迹）。元级别和基础级别的优化都在带有标准工具集的智能体化编码环境中执行： $$ \mathcal{T} = \{\texttt{Read}, \texttt{Write}, \texttt{Edit}, \texttt{Bash}, \texttt{Glob}, \texttt{Grep}, \texttt{TodoWrite}\} $$ **Meta-Harness** ([Lee et al. 2026](https://arxiv.org/abs/2603.28052)) moves another level deeper: the optimized object is the *code* that determines and optimizes what information should be stored, retrieved, and presented to the model. “Meta-” in its name means it is a harness for optimizing harnesses. **Meta-Harness** ([Lee 等 2026](https://arxiv.org/abs/2603.28052)) 则更深入了一层：其优化的对象是决定并优化哪些信息应该被存储、检索并呈现给模型的**代码**本身。其名称中的“Meta-（元）”意味着它是一个用于优化 Harness 的 Harness。

The Meta-Harness outer-loop optimization algorithm. (Image source: Lee et al. 2026)

Meta-Harness 外层循环优化算法。

The proposer for creating a new harness is itself a coding agent and the final output is a collection of harness candidates on the Pareto frontier. - The entire execution history is accessible via a file system, and thus the coding agent uses commands like `grep` or `cat` to read through it instead of shoveling everything into a single prompt context. - The proposed harness is a dictionary in the file system containing its own source code, scores, rollout trajectories, and state updates. - The mete-harness loop iteratively creates new harnesses, and only qualified ones are kept. 负责创建新 Harness 的提议者（proposer）本身也是一个编程智能体，最终输出的是位于帕累托前沿（Pareto frontier，指无法在不损害一个目标的情况下改善另一个目标的最优解集合）的候选 Harness 集合。 - 整个执行历史都可以通过文件系统访问，因此编程智能体会使用 `grep` 或 `cat` 等命令来读取历史记录，而不是将所有内容一股脑塞进单个提示词上下文中。 - 提议的 Harness 在文件系统中体现为一个目录，其中包含其自身的源代码、分数、执行轨迹以及状态更新。 - Meta-Harness 循环通过迭代不断创建新的 Harness，并且只保留合格的版本。

The performance of Meta-Harness on (Left) text classification with a small number of iterations and (Right) TerminalBench-2. Note that the search in the TerminalBench-2 experiment is initialized from Terminus-KIRA and Terminus-2, two very strong harnesses. (Image source: Lee et al. 2026)

Meta-Harness 在（左图）少量迭代的文本分类任务和（右图）TerminalBench-2 上的表现。请注意，TerminalBench-2 实验中的搜索是以 Terminus-KIRA 和 Terminus-2 这两个非常强大的 Harness 作为初始化起点的。

Still, the important lesson is clear: once harness design becomes an executable search space, a strong coding agent can exploit the same design space human engineers use. 尽管如此，一个重要的启示是明确的：一旦 Harness 设计成为了一个可执行的搜索空间，强大的编程智能体就能够像人类工程师一样探索和利用这个设计空间。 ### Workflow Design ### 工作流设计 Workflow design in harness engineering can be handcrafted by domain experts. Taking auto-research as an example, various frameworks have been proposed and tested. The **AI Scientist** system ([Lu et al. 2026](https://www.nature.com/articles/s41586-026-10265-5)) builds a pipeline to propose research ideas, write code, run experiments, analyze results, write a manuscript, and perform peer review. [Meng et al. (2026)](https://arxiv.org/abs/2605.26340) make verifiability the central design constraint in **ScientistOne**, where every claim (citation, numerical, methodological, conclusion) must trace to an evidence source and is audited by Chain-of-Evidence checks. Harness 工程中的工作流设计可由领域专家手工打造。以自动化研究（auto-research）为例，目前已提出并测试了各种框架。**AI Scientist** 系统 ([Lu 等 2026](https://www.nature.com/articles/s41586-026-10265-5)) 构建了一条流水线，用于提出研究想法、编写代码、运行实验、分析结果、撰写论文草稿并进行同行评审。[Meng 等 (2026)](https://arxiv.org/abs/2605.26340) 则将可验证性作为 **ScientistOne** 系统的核心设计约束，在该系统中，每一个声明（引用、数值、方法、结论）都必须能追溯到证据来源，并通过证据链（Chain-of-Evidence）进行审计。

AI Scientist pipeline for idea generation, experimentation, paper writing, and review. (Image source: Lu et al. 2026)

AI Scientist 从创意生成、实验、论文撰写到评审的流水线。

The **Autodata** agent ([Kulikov et al. 2026](https://arxiv.org/abs/2606.25996)) is designed to work as a data scientist for generating training and evaluation data. The main agent manages a *challenger* that proposes problems, a *weak solver*, a *strong solver*, and a *verifier/judge*, aiming to synthesize data at the “just right” level of difficulty, meaning that the strong solver succeeds but the weak solver fails. **Autodata** 智能体 ([Kulikov 等 2026](https://arxiv.org/abs/2606.25996)) 被设计为扮演数据科学家的角色，用于生成训练和评估数据。主智能体管理着一个提出问题的*挑战者（challenger）*、一个*弱求解器（weak solver）*、一个*强求解器（strong solver）*以及一个*验证者/裁判（verifier/judge）*，旨在合成难度“恰到好处”的数据——即强求解器能成功解决，而弱求解器会失败的数据。 In Autodata, the challenger prompt is updated iteratively according to feedback from the solvers and verifier. The limitation here is that synthesized tasks are used to fine-tune weak solvers but not strong solvers; if the loop cannot iteratively improve the strong model, it is more like indirect distillation over a generated prompt distribution, with less RSI flavor. 在 Autodata 中，挑战者的提示词会根据求解器和验证者的反馈进行迭代更新。其局限性在于，合成的任务仅用于微调弱求解器，而不是强求解器；如果这个循环不能迭代式地提升强模型，它就更像是在生成的提示词分布上进行间接的知识蒸馏，从而少了几分 RSI（递归自我提升）的意味。

Autodata agentic workflow design for generating synthetic training and evaluation data around challenger, solver, and verifier roles. (Image source: Kulikov et al. 2026)

围绕挑战者、求解器和验证者角色，用于生成合成训练和评估数据的 Autodata 智能体化工作流设计。

The design space for workflow is *enormous*, and naturally we can think of workflow design as a search problem, and therefore we should be able to find good solutions by algorithms rather than only manually craft them. Following this direction, **Automated Design of Agentic Systems** (ADAS; [Hu et al. 2025](https://arxiv.org/abs/2408.08435)) formulates agent design itself as an optimization problem, “meta-agent search” where a meta-agent proposes new designs of agentic workflows. 工作流的设计空间是**极其巨大**的。自然地，我们可以将工作流设计视为一个搜索问题，因此我们应该能够通过算法找到好的解决方案，而不是仅仅依靠手工打造。沿着这个方向，**自动化智能体系统设计（Automated Design of Agentic Systems, ADAS**; [Hu 等 2025](https://arxiv.org/abs/2408.08435)) 将智能体设计本身公式化为一个优化问题，即“元智能体搜索（meta-agent search）”，由一个元智能体来提出智能体化工作流的新设计。 1. Initialize an archive of agentic workflows with simple agents such as CoT and self-refine. 2. Ask a meta-agent to program new agents, all in *code*, inspired by existing solutions in the archive. - The meta-agent first generates a high-level description of the new workflow, and then implements it in code. - The draft program then goes through two self-refine steps (i.e. ask the model to provide feedback and then ask the same model to refine the previously generated outputs based on the feedback; [Madaan et al. 2023](https://arxiv.org/abs/2303.17651)) by the meta-agent to check its novelty. 3. Evaluate each new candidate and add successful ones back to the archive. 4. Repeat steps 2-3 until the maximum iteration count is reached. 1. 用简单的智能体（如思维链 CoT 和自我反思 self-refine）初始化一个智能体化工作流档案库。 2. 要求元智能体受到档案库中现有方案的启发，完全用**代码**来编程实现新的智能体。 - 元智能体首先生成新工作流的高层描述，然后用代码实现它。 - 草稿程序随后由元智能体进行两步自我改进（即要求模型提供反馈，然后让同一个模型根据反馈修改之前生成的输出；[Madaan 等 2023](https://arxiv.org/abs/2303.17651)），以检查其新颖性。 3. 评估每一个新的候选方案，并将成功的方案添加回档案库中。 4. 重复步骤 2-3，直到达到最大迭代次数。

Illustration of Automated Design of Agentic Systems (ADAS).
(Image source: Hu et al. 2025)

自动化智能体系统设计（ADAS）图解。

**AFlow** ([Zhang et al. 2025](https://arxiv.org/abs/2410.10762)) represents an agentic workflow as a graph, where nodes represent LLM-invoking actions and edges implement logical operations in code. The workflow optimization relies on [MCTS](https://en.wikipedia.org/wiki/Monte_Carlo_tree_search) (Monte Carlo Tree Search), a heuristic search algorithm famous for powering AI systems like AlphaGo by balancing exploration of new paths and exploitation of known good paths: **AFlow** ([Zhang 等 2025](https://arxiv.org/abs/2410.10762)) 将智能体化工作流表示为一个图，其中节点代表调用 LLM 的动作，边代表在代码中实现的逻辑操作。工作流优化依赖于 [MCTS（蒙特卡洛树搜索）](https://en.wikipedia.org/wiki/Monte_Carlo_tree_search)，这是一种著名的启发式搜索算法（曾助力 AlphaGo 等 AI 系统），它通过在探索新路径和利用已知好路径之间取得平衡来进行搜索： 1. Initialize the starting workflow $W_0$ in the tree with a template. 2. Select a workflow node using a soft mixture of score and uniform exploration. 3. Expand it by asking an LLM to produce a modified workflow conditioned on its evaluation performance. 4. Execute and evaluate the new workflow. 5. Add it back to the tree if the new workflow shows improvement within a budget of $N$ rounds. 6. Repeat steps 2-5 and stop when the top-$k$ average score plateaus or hit the budget. 1. 用一个模板在树中初始化起始工作流 $W_0$。 2. 结合得分和均匀探索的软混合策略，选择一个工作流节点。 3. 扩展该节点，即让 LLM 根据其评估表现生成一个修改后的工作流。 4. 执行并评估新工作流。 5. 如果新工作流在 $N$ 轮预算内表现出改进，则将其添加回搜索树中。 6. 重复步骤 2-5，当排名前 $k$ 的平均得分趋于平缓或耗尽预算时停止。

AFlow optimization process over a tree of workflow candidates. (Image source: Zhang et al. 2025)

AFlow 在候选工作流树上的优化过程。

Experiments of AFlow in QA, code, and math tasks showed decent improvement of AFlow over manually designed workflows and ADAS. AFlow 在问答（QA）、代码和数学任务中的实验表明，相较于手工设计的工作流和 ADAS，AFlow 取得了显著的改进。

AFlow experiments in comparison to manual methods and ADAS. (Image source: Zhang et al. 2025)

AFlow 与手工方法和 ADAS 的实验对比。

### Self-Improving Harness ### 自我提升的 Harness Either context engineering or workflow design is only one part of a harness. We need to search through the entire design space and optimize context-management logic, workflow, permissions, and many other harness components together. As we have seen in work like Meta-Harness, ADAS, and AFlow, **✨code✨** is a **universal language** for defining programs and systems. In simple words, a harness is code that programs how prompts, tool calls, subagents, control flow, memory, and workflow logic work together. If an LLM can optimize the code that executes agents, it can access a *much larger design space* than hand-written prompts. 无论是上下文工程还是工作流设计，都只是 Harness 的一部分。我们需要在整个设计空间中进行搜索，将上下文管理逻辑、工作流、权限以及许多其他 Harness 组件结合在一起进行优化。正如我们在 Meta-Harness、ADAS 和 AFlow 等工作中所看到的，**✨代码✨是定义程序和系统的通用语言**。简单来说，Harness 就是一段代码，它对提示词、工具调用、子智能体、控制流、记忆以及工作流逻辑如何协同工作进行编程。如果 LLM 能够优化执行智能体的这段代码，它就能触及比手工编写提示词**大得多**的设计空间。 **Self-Taught Optimizer** (STOP; [Zelikman et al. 2023](https://arxiv.org/abs/2310.02304)) is one of the early examples of recursive scaffolding improvement. A seed improver $I_0$ at step $t=0$ takes an initial solution $s$, a utility function $u$, and a black-box language model $M$, and returns an improved solution $s'$, that is, $s' = I(u, s; M)$. The goal of STOP is not directly to improve $s$ but *to improve the improver $I$ itself*. **自学习优化器（Self-Taught Optimizer, STOP**; [Zelikman 等 2023](https://arxiv.org/abs/2310.02304)）是递归框架优化的早期范例之一。在步骤 $t=0$ 时，一个种子优化器 $I_0$ 接收初始解决方案 $s$、效用函数 $u$ 以及黑盒语言模型 $M$，并返回改进后的解决方案 $s'$，即 $s' = I(u, s; M)$。STOP 的目标不是直接改进 $s$，而是**改进优化器 $I$ 本身**。 First, let’s define the meta-utility as the average utility of a given improver function $I$ over a collection of downstream tasks $\mathcal{D}$: 首先，我们将给定优化器函数 $I$ 在一组下游任务 $\mathcal{D}$ 上的平均效用定义为“元效用（meta-utility）”： $$ \hat{u}(I) \triangleq \frac{1}{|\mathcal{D}|} \mathbb{E}_{(u,s) \sim \mathcal{D}} [u(I(u,s; M))] $$ Because improving the improver function is an optimization problem itself, we can recursively get a new version of $I_t$ based on $I_{t-1}$’s performance measured by meta-utility via a self-improvement update: 由于改进优化器函数本身就是一个优化问题，我们可以通过自我提升的更新操作，根据元效用衡量的 $I_{t-1}$ 性能，递归地获得新版本的 $I_t$： $$ I_t = I_{t-1}(\hat{u}, I_{t-1}; M) $$

Algorithm of Self-Taught Optimizer (STOP). (Image source: Zelikman et al. 2023)

自学习优化器（STOP）算法。

In Zelikman et al. (2023)’s experiments, the improved improver discovered various strategies, such as genetic algorithms, decomposing and improving parts, multi-armed prompt bandits, simulated annealing, varying temperature, and beam/tree search. This is analogous to how a harness workflow can be represented as an object for optimization. 在 Zelikman 等人（2023）的实验中，不断进化的优化器发现了多种策略，例如遗传算法、分解并改进局部、多臂提示词老虎机（multi-armed prompt bandits）、模拟退火（这两种都是经典的优化探索算法）、调整温度参数以及束搜索/树搜索。这就如同将 Harness 工作流视作一个优化对象一样。

Examples of self-improvement strategies discovered by STOP. (Image source: Zelikman et al. 2023)

STOP 发现的自我提升策略示例。

A *cautionary* result in their findings is that STOP improved mean downstream performance across iterations with GPT-4 but degraded with weaker models like GPT-3.5 and Mixtral. Recursive structure alone is not enough. The base model must be *capable enough* to improve the mechanism. This implies that harness improvement enables better deployment of the model but intelligence is still the core. 在他们的发现中，一个值得**警惕**的结果是：STOP 在使用 GPT-4 时能在迭代中提升下游的平均性能，但如果换成 GPT-3.5 或 Mixtral 这样较弱的模型，性能反而会下降。仅靠递归结构是不够的。基础模型必须具备**足够的能力**去改进这一机制。这意味着，Harness 的改进能够实现模型更好的部署，但“智能”本身依然是核心。 A more recent work, **Self-Harness** ([Zhang et al. 2026](https://arxiv.org/abs/2606.09498)), relies on LLM agents to improve their own harness via a propose-evaluate-accept loop. 更近期的一项研究 **Self-Harness** ([Zhang 等 2026](https://arxiv.org/abs/2606.09498))，则依靠 LLM 智能体通过“提出-评估-接受”的循环来改进自身的 Harness。

Self-Harness uses a loop of weakness mining, bounded harness proposal, and validation to update a harness. (Image source: Zhang et al. 2026)

Self-Harness 使用弱点挖掘、有界 Harness 提议和验证的循环来更新 Harness。

The loop in Self-Harness has three stages: Self-Harness 的循环分为三个阶段： 1. *Weakness mining*: cluster failures into verifier-grounded failure patterns. - The current harness $h_t$ is used to evaluate on tasks and execution traces are collected for analysis. - Note that two runs can share the same verifier outcome in the error logs on the surface, such as timeout or missing artifact, while having different causal mechanisms. Therefore we need a failure record of rich information, containing the terminal verifier-level cause, the causal status of the relevant agent behavior, and the abstract agent mechanism exposed by the trace, to uncover the root causes. 1. *弱点挖掘（Weakness mining）*：将失败案例聚类为基于验证者的失败模式。 - 使用当前的 Harness $h_t$ 在任务上进行评估，并收集执行轨迹以供分析。 - 需要注意的是，表面上看，两次运行在错误日志中可能有相同的验证结果（例如超时或产出物丢失），但它们的因果机制可能截然不同。因此，我们需要一个包含丰富信息的失败记录——包括终端验证器层面的原因、相关智能体行为的因果状态，以及轨迹暴露出的抽象智能体机制——来揭示根本原因。 2. *Harness proposal*: propose bounded harness edits based on mined failure patterns. - The same model is invoked under $h_t$ as a proposer. - The model is provided with a bounded proposal context: (1) the editable surfaces of the current harness, (2) the verifier-grounded failure patterns from the evaluation system, (3) records of passing behaviors that should be preserved, and (4) summaries of previously attempted edits. - Harness edits should prefer recurrent error patterns that are addressable (e.g. not task-specific difficulty) and can be resolved by narrow changes. - Harness edit candidates should be distinct and diverse. 2. *Harness 提议（Harness proposal）*：根据挖掘出的失败模式，提出有界的 Harness 修改建议。 - 同一个模型在 $h_t$ 环境下被调用作为提议者。 - 为该模型提供一个有界的提议上下文：（1）当前 Harness 的可编辑部分，（2）来自评估系统的、基于验证者的失败模式，（3）应当保留的成功行为记录，以及（4）先前尝试修改的摘要。 - Harness 修改应该优先考虑那些经常发生且可解决的错误模式（例如并非针对特定任务的固有难度），且可通过小范围变更解决。 - Harness 候选修改方案应该具有区分度和多样性。 3. *Proposal validation*: validate and merge qualified edits to create a new harness $h_{t+1}$. - Candidate edits are evaluated by regression tests on held-in $D_{\text{in}}$ (for testing whether the weakness is resolved) and held-out $D_{\text{out}}$ (for checking whether other unknown issues were introduced) splits. - Candidates are accepted only if they have no regression on both held-in and held-out data. - Accepted candidates are merged to update the harness to $h_{t+1}$, while rejected candidates are logged without changing the active harness. 3. *提议验证（Proposal validation）*：验证并合并合格的修改，以创建一个新的 Harness $h_{t+1}$。 - 候选修改通过回归测试进行评估，测试集分为保留内（held-in）数据集 $D_{\text{in}}$（用于测试弱点是否被解决）和保留外（held-out）数据集 $D_{\text{out}}$（用于检查是否引入了其他未知问题）。 - 仅当候选方案在保留内和保留外数据上均未出现性能倒退（regression）时，才会被接受。 - 被接受的候选方案会被合并，将 Harness 更新为 $h_{t+1}$；而被拒绝的候选方案会被记录在案，但不改变当前处于激活状态的 Harness。 When running `MiniMax M2.5`, `Qwen3.5-35B-A3B`, and `GLM-5` on Terminal-Bench-2, Self-Harness was shown to learn model-specific harness instructions that target at different weaknesses of different base models and improve held-out pass rates. 在 Terminal-Bench-2 上运行 `MiniMax M2.5`、`Qwen3.5-35B-A3B` 和 `GLM-5` 时，Self-Harness 证明了其能够学习针对不同基础模型的弱点而量身定制的模型特异性 Harness 指令，并提升保留外测试集的通过率。 Self-harness type of work does raise my concerns that if a program is allowed to edit the OS system, abstraction boundaries are broken. The editable surface needs to be properly designed and the permission control and security layers need to live outside this loop. All the challenges around [reward hacking](https://lilianweng.github.io/posts/2024-11-28-reward-hacking/) still remain. Self-Harness 这类工作确实引起了我的担忧：如果允许一个程序编辑其“操作系统”，抽象边界就会被打破。可编辑的范围需要被妥善设计，权限控制和安全层必须独立于这个循环之外。围绕着[奖励破解（reward hacking）](https://lilianweng.github.io/posts/2024-11-28-reward-hacking/)的所有挑战依然存在。 ### Evolutionary Search ### 演化搜索 Evolutionary search is an optimization method inspired by natural selection (see my old post on [evolutionary algorithm](https://lilianweng.github.io/posts/2019-09-05-evolution-strategies/)). It evolves a population of solutions by mutating them and only keeping those with high “fitness” in the crowd (survival of the fittest). Evolutionary search comes in handy when (1) the search space is extensive or weirdly shaped; and (2) it is hard to optimize directly with gradients but easy to evaluate solutions. Harness search seems to be a good fit here. 演化搜索（Evolutionary search）是一种受自然选择启发的优化方法（详见我关于[演化算法](https://lilianweng.github.io/posts/2019-09-05-evolution-strategies/)的旧文）。它通过变异（mutating）解决方案的种群，并只保留群体中具有高“适应度”的方案（适者生存）来进行演化。在以下情况中，演化搜索会非常有用：（1）搜索空间广阔或形状奇特；且（2）很难直接使用梯度进行优化，但很容易评估解决方案的质量。Harness 搜索似乎非常适合这种方法。 Evolutionary search has been used in prompt engineering in the past studies. **Promptbreeder** ([Fernando et al. 2023](https://arxiv.org/abs/2309.16797)) optimizes task-specific prompts through a rich set of mutation operations, and interestingly the mutation prompts (i.e. instructions to an LLM to mutate a task prompt) are themselves also improved through evolution. **GEPA** ([Agrawal et al. 2025](https://arxiv.org/abs/2507.19457)) combines [reflection](https://lilianweng.github.io/posts/2023-06-23-agent/#self-reflection)-based prompting with evolutionary search and uses natural language reflection over trajectories of trial and error to propose prompt updates. 在以往的研究中，演化搜索已被应用于提示词工程。**Promptbreeder** ([Fernando 等 2023](https://arxiv.org/abs/2309.16797)) 通过一组丰富的变异操作来优化针对特定任务的提示词，有趣的是，变异提示词本身（即指示 LLM 变异任务提示词的指令）也通过演化得到了改进。**GEPA** ([Agrawal 等 2025](https://arxiv.org/abs/2507.19457)) 将基于[反思（reflection）](https://lilianweng.github.io/posts/2023-06-23-agent/#self-reflection)的提示技术与演化搜索相结合，使用自然语言反思试错轨迹，以此来提出提示词的更新建议。 [Novikov et al. (2025)](https://arxiv.org/abs/2506.13131) introduced **AlphaEvolve** as a coding-agent evolutionary search system, which stores a pool of candidate programs and prompts frozen LLMs to generate diffs for improvement. As the system repeatedly evaluates child programs and keeps successful ones, it discovers better solutions in time. [Novikov 等 (2025)](https://arxiv.org/abs/2506.13131) 提出了 **AlphaEvolve**，这是一个编程智能体演化搜索系统，它存储了一个候选程序池，并提示冻结状态（权重不更新）的 LLM 生成代码差异（diffs）来进行改进。随着系统反复评估子程序并保留成功的程序，它会逐渐发现更好的解决方案。

How AlphaEvolve works. (Image source: Novikov et al. 2025)

AlphaEvolve 是如何工作的。

A few details matter in the design of AlphaEvolve: - The prompt includes parent programs, results, instructions, and sometimes meta information. - The coding agent has access to the full repo, but code regions for improvement are explicitly marked with `# EVOLVE-BLOCK-START` and `# EVOLVE-BLOCK-END`. - Meta-prompt co-evolves with instructions and context as suggested by LLM, in a similar way as how we evolve solution programs. 在 AlphaEvolve 的设计中，有几个细节至关重要： - 提示词包含了父级程序、运行结果、指令，有时还有元信息。 - 编程智能体可以访问整个代码仓库，但需要改进的代码区域会通过 `# EVOLVE-BLOCK-START` 和 `# EVOLVE-BLOCK-END` 明确标记出来。 - 按照 LLM 的建议，元提示词（Meta-prompt）与指令和上下文共同演化，其方式与我们演化解决方案程序的方式相似。 Ablations show the evolution procedure, context in prompts, meta-prompts, full-file evolution and the use of stronger LLMs. 消融实验展示了演化过程、提示词中的上下文、元提示词、全文件演化以及使用更强 LLM 的重要作用。

Ablations show the value of everal designs in AlphaEvolve. (Image source: Novikov et al. 2025)

消融实验展示了 AlphaEvolve 中各项设计的价值。

Recent variants such as **ThetaEvolve** ([Wang et al. 2025](https://arxiv.org/abs/2511.23473)) combines evolutionary search with RL and in-context learning. **ShinkaEvolve** ([Lange et al. 2025](https://arxiv.org/abs/2509.19349)), on the other hand, introduced three new components to improve LLM sampling efficiency: 近期的变体研究，如 **ThetaEvolve** ([Wang 等 2025](https://arxiv.org/abs/2511.23473))，将演化搜索与强化学习（RL）和上下文学习（in-context learning）结合起来。而 **ShinkaEvolve** ([Lange 等 2025](https://arxiv.org/abs/2509.19349)) 则引入了三个新组件来提升 LLM 的采样效率： - More sample-efficient exploration by designing parent sampling to balance performance rank and offspring count. - Code-novelty rejection sampling by discarding candidates that are too similar to the existing population based on embedding-based cosine similarity. - Identifying good patterns in successful solutions in a meta-scratchpad to guide future mutation. - 通过设计平衡性能排名和后代数量的父级采样策略，实现更高效的样本探索。 - 代码新颖性拒绝采样：基于嵌入向量的余弦相似度（cosine similarity），丢弃那些与现有种群过于相似的候选方案。 - 在元暂存板（meta-scratchpad）中识别成功解决方案中的优秀模式，以此来指导未来的变异。 Unlike the methods above, which focus on solution improvement, **Darwin Gödel Machine** (DGM; [Zhang et al. 2025](https://arxiv.org/abs/2505.22954)) explicitly targets the evolution of an editable harness-code repository with an LLM-based coding agent. Precisely, this agent is allowed to modify its own harness. A follow-up work on Hyperagents ([Zhang et al. 2026](https://arxiv.org/abs/2603.19461)) introduced a meta-agent to control how to modify existing task agents to create new ones. 与上述专注于改进解决方案的方法不同，**达尔文·哥德尔机（Darwin Gödel Machine, DGM**; [Zhang 等 2025](https://arxiv.org/abs/2505.22954)）明确瞄准了使用基于 LLM 的编程智能体对可编辑的 Harness 代码库进行演化。准确地说，该智能体被允许修改其自身的 Harness。后续关于 Hyperagents（超智能体）的研究 ([Zhang 等 2026](https://arxiv.org/abs/2603.19461)) 引入了一个元智能体，用于控制如何修改现有任务智能体以创造新的智能体。 1. Start with one coding agent in the pool. 2. In each iteration, pick one parent with a probability proportional to its performance and inversely to the number of children it has, to modify and branch off to produce new agents. 3. The selected parent agent examines its own benchmark evaluation log and then proposes improvements to its own harness codebase to generate a new version of the coding agent. Code editing is implemented with two basic tools: (1) bash (args: ``) and (2) editor (args: `view/create/edit `). 4. New coding agents are evaluated, and only those with sufficiently high performance are added back into the pool. 5. Repeat steps 2-4 until some stop criteria hit. 1. 从池中的一个编程智能体开始。 2. 在每次迭代中，以正比于其性能、反比于其已有后代数量的概率，选择一个父代智能体，进行修改并拉出分支以产生新的智能体。 3. 被选中的父代智能体检查自身的基准评估日志，然后向自身的 Harness 代码库提出改进建议，以生成新版本的编程智能体。代码编辑通过两个基础工具实现：（1）bash（参数：``）和（2）editor（参数：`view/create/edit `）。 4. 评估新的编程智能体，只有性能足够高的才会被加回池中。 5. 重复步骤 2-4，直到满足某种停止条件。 DGM is harness evolution under a fixed model. In experiments with `Claude 3.5 Sonnet` as the base LLM and simple initial harness configs, the DGM-discovered agents are comparable to or outperform handcrafted agents on SWE-bench Verified (20% to 50%) and Polyglot (14.2% to 30.7%). DGM 是在固定模型下的 Harness 演化。在以 `Claude 3.5 Sonnet` 为基础 LLM 和简单初始 Harness 配置的实验中，DGM 发现的智能体在 SWE-bench Verified（从 20% 提升到 50%）和 Polyglot（从 14.2% 提升到 30.7%）测试集上，媲美甚至超越了人工精心打造的智能体。 This family of methods works well when candidate solutions are automatically evaluable and candidate fitness is easy to quantify, such as matrix multiplication, GPU kernel optimization, algorithm contests, datacenter scheduling. It struggles with domains where evaluation is slow, ambiguous, or mostly heuristic-based. The compute efficiency and effectiveness of evolution are also concerns. 当候选解决方案可以被自动评估，且候选适应度（fitness）易于量化时（如矩阵乘法、GPU内核优化、算法竞赛、数据中心调度），这类方法效果显著。但它在评估缓慢、模糊或主要基于启发式规则的领域中则表现不佳。计算效率和演化的有效性也是需要考虑的担忧点。 ### Joint Optimization with Model Weights ### 与模型权重的联合优化 Harness evolution changes the non-parametric system around the model. To enable full self-improvement, the model can totally be allowed to update its own weights at the same time. The weight update can be implemented via improvements in the model training pipeline or continual learning at test time. The topic of continual learning is worthy of its own post in the future. Harness 演化改变的是模型外围的非参数系统。为了实现完全的自我提升，完全可以允许模型同时更新其自身的参数权重。权重更新可以通过改进模型训练流水线，或在测试时进行持续学习（continual learning）来实现。持续学习这个话题非常值得以后专门写一篇文章探讨。 **SIA** ([Hebbar et al. 2026](https://arxiv.org/abs/2605.27276)) is an early attempt to combine harness improvement and model-parameter updates in the same optimization loop, with three components in the design: - *Meta-Agent*: proposes the initial harness. - *Task-Specific Agent*: executes the task. - *Feedback-Agent*: chooses whether to update the harness or the model weights based on recent trajectories. **SIA** ([Hebbar 等 2026](https://arxiv.org/abs/2605.27276)) 是将 Harness 改进和模型参数更新结合在同一个优化循环中的早期尝试，其设计包含三个组件： - *元智能体（Meta-Agent）*：提出初始的 Harness。 - *任务特定智能体（Task-Specific Agent）*：执行任务。 - *反馈智能体（Feedback-Agent）*：根据近期的执行轨迹，选择是更新 Harness 还是更新模型权重。

The Feedback-Agent in SIA decides the next iteration type. (Image source: Hebbar et al. 2026)

SIA 中的反馈智能体决定下一次迭代的类型。

There are a few confounding choices in SIA’s experiments that make the results hard to interpret. For example, the task-specific agent is much weaker than the models used for the Meta-Agent and Feedback-Agent (`gpt-oss-120b` vs `Claude Sonnet 4.6`), and the baselines are too weak to cross-reference cleanly against related methods. I would consider the direction interesting, but the evidence provisional. Yet many challenges, such as training stability and Goodhart effect, still remain open. SIA 的实验中存在一些容易混淆的选择，使得结果难以解释。例如，任务特定智能体使用的模型比元智能体和反馈智能体使用的模型弱得多（`gpt-oss-120b` 相比于 `Claude Sonnet 4.6`），并且作为对比的基线方法太弱，无法与其他相关方法进行清晰的交叉参考。我认为这个方向很有趣，但现有证据仍是初步的。诸如训练稳定性和古德哈特效应（Goodhart effect，即当一个指标被用作目标时，它就不再是一个好指标）等许多挑战依然悬而未决。 ## Future Challenges ## 未来的挑战 The AI Scientist line of work is a strong demonstration that an expert-designed harness can coordinate a large portion of auto-research loop, experimented in the form of writing research papers. But paper production is not identical to scientific discovery. A system can write a plausible manuscript while still having fabricated citations, implementation drift, or weak experimental results. AI Scientist 系列工作有力地证明了，专家设计的 Harness 能够协调自动研究循环中的很大一部分工作（以撰写研究论文的形式进行了实验）。但是，论文产出并不等同于科学发现。一个系统可能写出一篇看似合理的初稿，但其中依然包含虚构的引用、实现偏差或薄弱的实验结果。 [Trehan & Chopra (2026)](https://arxiv.org/abs/2601.03315) tested whether LLMs can go from a research idea to a paper with minimal scaffolding and basic tools (i.e., `read_file`, `write_file`, `llm_search`, `list_files`). Each idea had a dedicated workspace where agents could generate and read documents as part of context. They experimented in three domains (world models, multi-agent RL, AI safety & alignment), with each domain containing 45-50 high-quality seed documents to inspire new ideas. Only four ideas were selected by human experts to run through the full pipeline, and only one was fully executed into a paper. They observed six recurring failure modes in the experiments: [Trehan 和 Chopra (2026)](https://arxiv.org/abs/2601.03315) 测试了 LLM 是否能在最少的支架辅导和基础工具（即 `read_file`, `write_file`, `llm_search`, `list_files`）的帮助下，完成从研究灵感到写成论文的全过程。每个想法都有一个专属的工作区，智能体可以在那里生成和阅读文档作为上下文的一部分。他们在三个领域（世界模型、多智能体强化学习、AI安全与对齐）进行了实验，每个领域提供 45-50 篇高质量种子文档以激发新想法。最终只有四个想法被人类专家选中运行完整的流水线，而只有一个被完整执行并写成论文。他们在实验中观察到了六种反复出现的失败模式： - *Bias toward training-data defaults*: use old libraries, stale commands, standard formats, or assumptions not grounded in the actual repository or dataset. - *Implementation drift under execution pressure*: when implementation becomes technically complex, the model may move toward a common simpler solution rather than the proposed method. - *Memory and context degradation*: long-horizon projects lose critical details unless logs are written as persistent artifacts. - *Over-optimism*: the model declares success despite noisy or failed experiments, similarly observed as “p-hacking and eureka-ing” pattern by [Bubeck et al. (2025)](https://arxiv.org/abs/2511.16072) where models can introduce “numerical duct tape” and declare victory when signals are still noise. - *Insufficient domain intelligence*: the model lacks tacit craft knowledge, e.g. predicting implementation complexity, judging whether an experimental result is plausible, or knowing which baselines matter. - *Weak scientific taste*: experiments may be executable but fail to answer the right question. - *偏向训练数据默认值的偏见（Bias toward training-data defaults）*：使用旧的代码库、陈旧的命令、标准格式，或者做出并未基于实际代码仓库或数据集的假设。 - *执行压力下的实现漂移（Implementation drift under execution pressure）*：当代码实现变得技术复杂时，模型可能会退缩至常见的简单解决方案，而不是坚持所提议的方法。 - *记忆和上下文退化（Memory and context degradation）*：除非日志被记录为持久化产出物，否则长周期的项目会丢失关键细节。 - *过度乐观（Over-optimism）*：尽管实验充满了噪音或根本就是失败的，模型依然宣布成功。这与 [Bubeck 等 (2025)](https://arxiv.org/abs/2511.16072) 观察到的“p-hacking（P值操纵，指不断摆弄数据直至出现统计显著性）与 eureka-ing（虚假顿悟）”模式类似，即模型可能会引入“数字胶带（缝合怪代码）”并在信号依然是噪音时就宣称胜利。 - *领域智能不足（Insufficient domain intelligence）*：模型缺乏难以言传的行业经验（tacit craft knowledge），例如预测实现复杂度、判断实验结果是否合理，或者知道哪些基准测试才是真正重要的。 - *科学品味低下（Weak scientific taste）*：实验可能成功执行了，但未能回答正确的研究问题。 Toward full RSI, researchers have made real progress, but several bottlenecks remain. 在通往完全 RSI（递归自我提升）的道路上，研究人员已取得了实质性的进展，但仍存在几个瓶颈。 **1. Weak and fuzzy evaluators.** Many research claims do not have a fast and precise verifier, and the same is true for many real-world tasks. Current self-improvement loops work best for tasks when evaluation metrics are measurable and objective, similar as [how RL works](https://lilianweng.github.io/posts/2018-02-19-rl-overview/). **1. 评估器薄弱且模糊。** 许多研究声明并没有快速且精确的验证工具，现实世界中的许多任务也是如此。当前的自我提升循环在评估指标可测量且客观的任务上表现最佳，这与[强化学习（RL）的工作原理](https://lilianweng.github.io/posts/2018-02-19-rl-overview/)类似。 Research taste, novelty, and long-term scientific value are much harder to measure. For example, research taste often mixes problem framing, experimental design, and judgment about which surprising results are worth pursuing and which failure cases are worth retries. 研究品味、新颖性以及长期科学价值要难以衡量得多。例如，研究品味通常融合了问题框架的构建、实验设计，以及对哪些令人惊讶的结果值得深究、哪些失败案例值得重试的判断。 **2. Context and memory lifecycle.** Memory grows as AI agents become more autonomous and independent. A useful harness needs to manage context and memory to complement existing limitation in long-context generation while still maximizing the success of long-horizon tasks. Since humans are able to maintain memory through our life time, I see an anoloy here that [context engineering](#context-engineering) will and should become a core part of intelligence, rather than staying in the software system layer. **2. 上下文和记忆的生命周期。** 随着 AI 智能体变得越来越自主和独立，其记忆也在不断增长。一个好用的 Harness 需要妥善管理上下文和记忆，以弥补当前模型在长上下文生成中的局限，同时最大化长周期任务的成功率。既然人类能够在一生中维持记忆，我在这里看到了一个类比：[上下文工程](#context-engineering)将不可避免地（也理应）成为智能核心的一部分，而不是仅仅停留在软件系统层面。 **3. Negative results.** Researchers are incentivized to publish successful results and thus literature is biased toward successes. LLMs trained on a vast amount of data (mostly human created, at least for now, lol) may be bad at deciding when to abandon a hypothesis, report a negative result, or even acknowledge a failure due to the imablance of success vs failure cases in data. A research harness should make failed attempts easy to preserve, as learning from failure is the best way to trim down the task search space. **3. 负面结果。** 研究人员在激励机制下更倾向于发表成功的结果，因此文献中充满了“幸存者偏差”。LLM 训练于海量数据之上（至少目前大部分还是人类创造的，笑），由于数据中成功与失败案例的失衡，它们可能不善于决定何时放弃一个假设、报告一个负面结果，甚至不愿承认失败。一个研究 Harness 应该使失败的尝试易于保存，因为从失败中学习是缩小任务搜索空间的最佳途径。 **4. Diversity collapse.** Evolutionary and RL loops tend to exploit known high-reward patterns. We need [mechanisms](https://lilianweng.github.io/posts/2020-06-07-exploration-drl/) to prevent the population from collapsing into variants of the same solution. This is especially critical for open-ended research, where the best path may initially look worse under the current evaluator. **4. 多样性崩溃。** 演化和 RL 循环倾向于不断利用已知的、高回报的模式。我们需要相关[机制](https://lilianweng.github.io/posts/2020-06-07-exploration-drl/)来防止种群坍缩为同一个解决方案的各种变体。这对于开放式研究尤为关键，因为最佳路径在当前的评估器下，最初看起来可能会更糟。 **5. [Reward hacking](https://lilianweng.github.io/posts/2024-11-28-reward-hacking/).** A self-improvement loop optimizes whatever signal it is given. This is a classic manifestation of Goodhart's Law—when a measure becomes a target, it ceases to be a good measure. If the reward comes from unit tests, the agent may overfit to tests; if it comes from a judge model, it may learn reward hacking tricks specific to this judge; if it comes from benchmark scores, it may exploit benchmark artifacts. **5. [奖励破解（Reward hacking）](https://lilianweng.github.io/posts/2024-11-28-reward-hacking/)。** 自我提升循环会优化任何给予它的信号。这是古德哈特定律（Goodhart's Law）的经典体现——当一个衡量指标成为目标时，它就不再是一个好指标。如果奖励来自单元测试，智能体可能会对测试集过拟合；如果奖励来自裁判模型（judge model），它可能会学习专门针对该裁判的欺骗技巧；如果奖励来自基准测试分数，它可能会利用基准测试的漏洞。 The evaluator and permission control should likely sit outside the loop that evolves harness, with held-out tests, trace audits, and human review at decision points that matter—how much oversight can be scaled up and automated remains an open research area. 评估器和权限控制或许应该置于 Harness 演化循环之外，在关键的决策点引入保留外（held-out）测试、轨迹审计和人工审查——如何扩大这种监督的规模并将其自动化，仍是一个开放的研究领域。 **6. Long-term success.** An extrinsic loop of optimization works on rewards outside of individual rollouts that we can simulate in training sandbox. **6. 长期成功。** 外在的优化循环作用于单个执行轨迹之外的奖励上，而这些是我们在训练沙盒中能够模拟的。 Take coding agent as an example. Coding agents have already increased daily productivity in software engineering, but many optimization goals are still too short-term. It can often complete the task at hand, but less obvious how it should protect the long-term health of a repo collectively maintained by hundreds or thousands of engineers. Standard sandbox-based RLVR-style training rarely captures maintainability, ownership boundaries, migration cost, backwards compatibility, or future debugging burden. 以编程智能体为例。编程智能体已经提高了软件工程的日常生产力，但许多优化目标仍然过于短视。它通常能够完成手头的任务，但如何保护由成百上千名工程师共同维护的代码仓库的长期健康，却并非显而易见。标准的基于沙盒的 RLVR（基于规则验证的强化学习）式训练很少能捕捉到代码可维护性、所有权边界、迁移成本、向后兼容性或未来的调试负担。 **7. The role of humans.** Humans should move up the stack, not be removed from the loop, meaning that human should provide oversight at the right time, at the right abstraction level and our system design should consider when and how to set up such touch points. **7. 人类的角色。** 人类应该向技术栈的更高层移动，而不是被从整个闭环中移除。这意味着人类应该在正确的时间、正确的抽象层级上提供监督，而我们的系统设计应当考虑到在何时、以何种方式设置这些人工触点（touch points）。 Many challenges listed above need human’s feedback and steering. After all, we are building the technology for better future of humanity, not other way around. 上面列出的许多挑战都需要人类的反馈和引导。毕竟，我们开发技术是为了人类更美好的未来，而不是相反。 ## Citation ## 引用 Please cite this work as: > Weng, Lilian. “Harness Engineering for Self-Improvement”. Lil’Log (Jul 2026). https://lilianweng.github.io/posts/2026-07-04-harness/ Or use the BibTeX citation: ```properties @article{weng2026harness, title = {Harness Engineering for Self-Improvement}, author = {Weng, Lilian}, journal = {lilianweng.github.io}, year = {2026}, month = {July}, url = "https://lilianweng.github.io/posts/2026-07-04-harness/" } ``` ## Appendix: Some useful benchmarks ## 附录：一些有用的基准测试（Benchmarks） - **[PaperBench](https://arxiv.org/abs/2504.01848)**: replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. - Each replication task is decomposed into smaller, individually gradable tasks. - 8,316 rubrics in total, co-developed with the paper authors. - The best model at the time (`Claude 3.5 Sonnet`, ~21%) does not outperform ML PhDs. - Includes PaperBench, PaperBench Code-Dev (a lighter version), and JudgeEval. - **[PaperBench](https://arxiv.org/abs/2504.01848)**：从零开始复现 20 篇 ICML 2024 Spotlight 和 Oral 论文，包括理解论文贡献、开发代码库并成功执行实验。 - 每个复现任务被分解为更小的、可单独评分的任务。 - 共有 8,316 个评分标准（rubrics），与论文原作者共同开发。 - 当时最好的模型（`Claude 3.5 Sonnet`，约 21%）表现未能超过机器学习领域的博士生。 - 包含 PaperBench、PaperBench Code-Dev（轻量版）和 JudgeEval。 - **[CORE-Bench](https://arxiv.org/abs/2409.11363)**: evaluate computational reproducibility of published research. - 270 tasks based on 90 scientific papers across computer science, social science, and medicine. - Tasks involve reproducing results from provided code and data. - Includes multiple difficulty levels and both language-only and vision-language tasks. - The best reported agent at the time (`GPT-4o` and `GPT-4o-mini`) achieved only 21% accuracy on the hardest task. - **[CORE-Bench](https://arxiv.org/abs/2409.11363)**：评估已发表研究的计算可复现性。 - 基于计算机科学、社会科学和医学领域的 90 篇科学论文构建了 270 个任务。 - 任务涉及从提供的代码和数据中复现结果。 - 包含多个难度级别，既有纯语言任务，也有视觉-语言任务。 - 当时报告中最好的智能体（`GPT-4o` 和 `GPT-4o-mini`）在最难任务上的准确率仅为 21%。 - **[ScienceAgentBench](https://arxiv.org/abs/2410.05080)**: evaluate LLM agents for data-driven scientific discovery. - Extracts 102 tasks from 44 peer-reviewed publications in four disciplines (math, chemistry, biology, geography). - Covers basic data-science tasks in these domains: data processing, model development, data analysis, and information visualization. - **[ScienceAgentBench](https://arxiv.org/abs/2410.05080)**：评估用于数据驱动科学发现的 LLM 智能体。 - 从四个学科（数学、化学、生物学、地理学）的 44 篇同行评审出版物中提取了 102 个任务。 - 涵盖了这些领域的基础数据科学任务：数据处理、模型开发、数据分析和信息可视化。 - **[RE-Bench](https://arxiv.org/abs/2411.15114)**: evaluate frontier AI agents on realistic ML research-engineering envs against human experts. - 7 challenging, open-ended ML research-engineering environments. - Each environment = (scoring function, starting solution, reference solution); each can be run with 8 or fewer H100 GPUs. - Examples: optimize a kernel, run a scaling-law experiment, fix an embedding, fine-tune GPT-2 for QA, etc. - Includes data from 71 eight-hour attempts by 61 distinct human experts. - Human experts achieved non-zero score in 82% of 8-hour attempts; 24% matched or exceeded strong reference solutions. - Best AI agents scored 4× higher than humans at a 2-hour budget, but humans had better returns to longer budgets and exceeded agents at 8-hour and 32-hour settings. - **[RE-Bench](https://arxiv.org/abs/2411.15114)**：在逼真的机器学习研究工程环境中，评估前沿 AI 智能体与人类专家的对比。 - 7 个具有挑战性的、开放式机器学习研究工程环境。 - 每个环境 =（评分函数，起始方案，参考方案）；每个环境都可以在 8 张或更少的 H100 GPU 上运行。 - 示例：优化一个 GPU kernel，运行缩放定律（scaling-law）实验，修复一个嵌入（embedding），为 QA 问答微调 GPT-2 等。 - 包含 61 位不同人类专家进行的 71 次 8 小时尝试的数据。 - 人类专家在 82% 的 8 小时尝试中获得了非零分数；24% 匹配或超过了强大的参考方案。 - 在 2 小时预算下，最好的 AI 智能体得分比人类高 4 倍，但人类在更长预算下有更好的回报，并在 8 小时和 32 小时设置下超过了智能体。 - **[MLE-bench](https://arxiv.org/abs/2410.07095)**: evaluate ML engineering agents on offline Kaggle competitions. - Contains 75 ML-engineering competitions curated from Kaggle. - Tests training models, preparing datasets, running experiments, and submitting predictions to grading scripts. - Uses Kaggle public leaderboards as human baselines. - Best setup in the paper, `o1-preview` with AIDE scaffolding, reached at least Kaggle bronze-medal level in 16.9% of competitions. - Includes resource-scaling and contamination analyses. - **[MLE-bench](https://arxiv.org/abs/2410.07095)**：在离线 Kaggle 竞赛上评估机器学习工程智能体。 - 包含从 Kaggle 精选出的 75 个机器学习工程竞赛。 - 测试模型训练、数据集准备、实验运行以及向评分脚本提交预测的能力。 - 使用 Kaggle 公开排行榜作为人类基线。 - 论文中最好的设置——使用 AIDE 脚手架的 `o1-preview`，在 16.9% 的竞赛中达到了至少 Kaggle 铜牌水平。 - 包含资源扩展和数据污染（contamination）分析。 - **[KernelBench](https://arxiv.org/abs/2502.10517)**: evaluate correctness and speed for generated GPU kernels. - 250 PyTorch tasks to evaluate whether LLM can write fast and correct kernels. - The evaluation metric fast_p = the percentage of generated kernels that are correct and faster than baseline. - **[KernelBench](https://arxiv.org/abs/2502.10517)**：评估生成的 GPU kernel 的正确性和速度。 - 250 个 PyTorch 任务，用于评估 LLM 是否能写出快速且正确的内核。 - 评估指标 fast_p = 生成内核既正确又比基线更快的百分比。 ## References ## 参考文献 *(References are kept in their original format for standard citation purposes)* *（参考文献保持原文格式以符合学术引用标准）* [1] Good, I. J. ["Speculations Concerning the First Ultraintelligent Machine."](https://philpapers.org/rec/GOOSCT) *Advances in Computers*, 6:31–88, 1965. [2] Yudkowsky, Eliezer. ["Recursive Self-Improvement."](https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement) LessWrong, 2008. [3] Choi, et al. ["Anchored Self-Play for Code Repair."](https://openreview.net/forum?id=lTbBFAoPSA) ICML 2026. [4] Zhao, et al. ["Absolute Zero: Reinforced Self-play Reasoning with Zero Data."](https://arxiv.org/abs/2505.03335) arXiv preprint arXiv:2505.03335, 2025. [5] Yuan, et al. ["Self-Rewarding Language Models."](https://arxiv.org/abs/2401.10020) arXiv preprint arXiv:2401.10020, 2024. [6] Chen, et al. ["Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models."](https://arxiv.org/abs/2401.01335) ICML 2024. [7] Zhang, et al. ["Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models."](https://arxiv.org/abs/2510.04618) ICLR 2026. [8] Ye, et al. ["Meta Context Engineering via Agentic Skill Evolution."](https://arxiv.org/abs/2601.21557) arXiv preprint arXiv:2601.21557, 2026. [9] Lee, et al. ["Meta-Harness: End-to-End Optimization of Model Harnesses."](https://arxiv.org/abs/2603.28052) arXiv preprint arXiv:2603.28052, 2026. [10] Lu, et al. ["Towards end-to-end automation of AI research."](https://www.nature.com/articles/s41586-026-10265-5) *Nature*, 651:914–919, 2026. [11] Meng, et al. ["ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence."](https://arxiv.org/abs/2605.26340) arXiv preprint arXiv:2605.26340, 2026. [12] Kulikov, et al. ["Autodata: An agentic data scientist to create high quality synthetic data."](https://arxiv.org/abs/2606.25996) arXiv preprint arXiv:2606.25996, 2026. [13] Hu, Lu, and Clune. ["Automated Design of Agentic Systems."](https://arxiv.org/abs/2408.08435) ICLR 2025. [14] Madaan, et al. ["Self-Refine: Iterative Refinement with Self-Feedback."](https://arxiv.org/abs/2303.17651) NeurIPS 2023. [15] Zhang, et al. ["AFlow: Automating Agentic Workflow Generation."](https://arxiv.org/abs/2410.10762) ICLR 2025. [16] Zelikman, et al. ["Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation."](https://arxiv.org/abs/2310.02304) COLM 2024. [17] Zhang, et al. ["Self-Harness: Harnesses That Improve Themselves."](https://arxiv.org/abs/2606.09498) arXiv preprint arXiv:2606.09498, 2026. [18] Fernando, et al. ["Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution."](https://arxiv.org/abs/2309.16797) arXiv preprint arXiv:2309.16797, 2023. [19] Agrawal, A. et al. ["GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning."](https://arxiv.org/abs/2507.19457) arXiv preprint arXiv:2507.19457, 2025. [20] Novikov, et al. ["AlphaEvolve: A coding agent for scientific and algorithmic discovery."](https://arxiv.org/abs/2506.13131) arXiv preprint arXiv:2506.13131, 2025. [21] Lange, Imajuku, and Cetin. ["ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution."](https://arxiv.org/abs/2509.19349) arXiv preprint arXiv:2509.19349, 2025. [22] Wang, et al. ["ThetaEvolve: Test-time Learning on Open Problems."](https://arxiv.org/abs/2511.23473) arXiv preprint arXiv:2511.23473, 2025. [23] Zhang, et al. ["Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents."](https://arxiv.org/abs/2505.22954) arXiv preprint arXiv:2505.22954, 2025. [24] Zhang, et al. ["Hyperagents."](https://arxiv.org/abs/2603.19461) arXiv preprint arXiv:2603.19461, 2026. [25] Yuksekgonul, et al. ["Learning to Discover at Test Time."](https://arxiv.org/abs/2601.16175) arXiv preprint arXiv:2601.16175, 2026. [26] Riaz, et al. ["Epistemic Uncertainty for Test-Time Discovery."](https://arxiv.org/abs/2605.11328) arXiv preprint arXiv:2605.11328, 2026. [27] Hebbar, et al. ["SIA: Self Improving AI with Harness & Weight Updates."](https://arxiv.org/abs/2605.27276) arXiv preprint arXiv:2605.27276, 2026. [28] Trehan and Chopra. ["Why LLMs Aren’t Scientists Yet: Lessons from Four Autonomous Research Attempts."](https://arxiv.org/abs/2601.03315) arXiv preprint arXiv:2601.03315, 2026. [29] Bubeck, et al. ["Early science acceleration experiments with GPT-5."](https://arxiv.org/abs/2511.16072) arXiv preprint arXiv:2511.16072, 2025. [30] Starace, et al. ["PaperBench: Evaluating AI’s Ability to Replicate AI Research."](https://arxiv.org/abs/2504.01848) ICML 2025. [31] Wijk, et al. ["RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts."](https://arxiv.org/abs/2411.15114) ICML 2025. [32] Chan, et al. ["MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering."](https://arxiv.org/abs/2410.07095) arXiv preprint arXiv:2410.07095, 2024. [33] Chen, et al. ["ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery."](https://arxiv.org/abs/2410.05080) ICLR 2025. [34] Siegel, et al. ["CORE-Bench: Fostering the Credibility of Published Research Through a Computational Reproducibility Agent Benchmark."](https://arxiv.org/abs/2409.11363) TMLR 2024. [35] Ouyang, et al. ["KernelBench: Can LLMs Write Efficient GPU Kernels?"](https://arxiv.org/abs/2502.10517) arXiv preprint arXiv:2502.10517, 2025.