这个项目的分工很直白：用 ChatGPT 做规划脑，保留 Codex 的 harness 做执行^[21]。它要解决的是单一工具的两难：对话类模型擅长澄清需求与拆解方案，但不擅长在真实仓库里持续执行；编码 harness 恰恰相反。做法上通过 MCP 把两边连接起来，让规划与执行各自发挥。值得借鉴的是把「想」与「做」拆到两个系统，并用协议而不是复制粘贴来交接；风险是跨系统交接会引入状态不同步与权限扩散问题，需要明确哪一侧持有真实状态。

The division of labour is explicit: use ChatGPT as the planning brain while keeping the Codex harness for execution^[21]. It addresses the dilemma of any single tool: chat models are good at clarifying requirements and decomposing plans but weak at sustained execution in a real repository, while coding harnesses are the reverse. MCP connects the two so each does its part. Worth borrowing is splitting thinking from doing across two systems and handing off by protocol rather than copy-paste; the risk is state desynchronisation and permission sprawl across the boundary, so decide which side owns the truth.

这个仓库提供免费的 SEO MCP 服务与开源的 SEO/GEO 技能，覆盖关键词研究、排名跟踪、审计、外链以及基于真实 GSC/GA4/广告数据的 AI 可见度分析^[22]。它要解决的是垂直领域知识难以接入 Agent 的问题：SEO 的判断依赖平台数据与时序变化，单纯写提示词无法覆盖。做法上把领域流程封装成技能与工具，让 Agent 直接调用真实数据源。值得借鉴的是「领域数据 + 领域流程 = 可安装技能」这个打包方式；风险是这类技能强依赖第三方数据源的政策与配额，数据源变更时整包失效。

This repo offers a free SEO MCP server plus open SEO/GEO skills covering keyword research, rank tracking, audits, backlinks and AI-visibility analysis on real GSC/GA4/ads data^[22]. It addresses the difficulty of wiring vertical expertise into agents: SEO judgements depend on platform data and time series, which prompts alone cannot cover. Domain workflows are packaged as skills and tools so the agent calls real data sources directly. Worth borrowing is the packaging pattern of domain data plus domain workflow equals an installable skill; the risk is heavy dependence on third-party data-source policy and quotas, where a source change breaks the whole pack.

这个项目是面向 Agent 的长期记忆运行时：以纯 Markdown 作为唯一真源，本地排序检索，并有一层独立的「睡眠期」管理进程负责整理，Claude Code 与 Codex 共用同一份存储，不需要 API key^[23]。它要解决的是记忆系统的可持续性：写入容易，但过期、重复与冲突的清理长期没人负责，最终记忆库变成噪声池。做法上把「整理」与「使用」分成两个进程，让整理可以离线慢慢做。值得借鉴的是把记忆维护设计成定期任务而不是在线副作用；风险是整理策略出错会静默篡改历史，需要可回滚与人工抽检。

This project is a long-term memory runtime for agents: plain Markdown as the source of truth, local ranked retrieval, and an independent sleep-time Manage layer, with Claude Code and Codex sharing one store and no API key required^[23]. It addresses memory sustainability: writes are easy, but expiring, deduplicating and resolving conflicting entries has no owner, so the store degrades into noise. The design separates managing from using so maintenance can run offline. Worth borrowing is treating memory maintenance as a scheduled job rather than an online side effect; the risk is that a faulty policy silently rewrites history, so rollback and sampling are needed.

这个项目是 x64dbg 的原生 MCP（Model Context Protocol）插件，把调试器完整能力通过 HTTP 暴露给 AI 助手：设断点、单步、读内存、导出寄存器等，用 Zig 写成零依赖单文件^[24]。它要解决的是逆向与恶意软件分析的效率问题：分析者大量时间花在重复的机械操作上，而这些操作恰好适合 Agent 执行。做法上不重写调试器，而是给它加一层标准协议接口。值得借鉴的是用协议层为既有成熟工具接上 Agent，而不是重做一个简化版；风险是让模型直接驱动调试器意味着极高权限，必须在隔离环境与快照中使用。

This project is a native MCP (Model Context Protocol) plugin for x64dbg exposing the debugger's full functionality over HTTP to any compatible AI assistant: set breakpoints, step, read memory, dump registers, written in Zig as a zero-dependency single binary^[24]. It addresses efficiency in reverse engineering and malware analysis, where analysts burn time on mechanical operations that agents handle well. Rather than rewriting a debugger, it adds a standard protocol interface. Worth borrowing is attaching agents to mature tools through a protocol layer instead of rebuilding simplified versions; the risk is that model-driven debugging is extremely privileged, so use isolated environments and snapshots.

headcount 的设计相当特别：把 Agent 组织成一家公司——15 个以上部门、125 个以上技能，每个都可独立安装，并要求引用能定案的标准与监管条款^[25]，可运行在 Claude Code 与 ChatGPT 上。它要解决的是通用 Agent 缺少「职责与边界」的问题：一个万能助手很难承担需要分工、复核与专业依据的工作。做法上用组织结构定义分工与升级路径，用标准引用定义判断依据。值得借鉴的是把「谁在什么情况下必须复核」写进组织设定；风险是组织结构一旦固化反而会限制跨领域任务，且技能质量差异会直接传导到输出。

headcount has an unusual design: an agent organisation structured as a company — 15+ departments, 125+ independently installable skills, each citing the standards and regulators that settle the question^[25], running in Claude Code and ChatGPT. It addresses the missing responsibility and boundaries of a general-purpose agent: a single do-everything assistant struggles with work needing division of labour, review and professional basis. Organisational structure defines roles and escalation paths, while standard citations define the basis for judgement. Worth borrowing is writing “who must review what, under which conditions” into the org definition; the risk is rigidity plus variance in skill quality propagating to output.

reverify 的口号是「别再让你的 AI 编东西：它提议，确定性工具裁决，每个论断都对真值核查并附证据」，且已核实的事实与上下文能跨越会话重置保留^[26]。它要解决的是幻觉在长任务里的累积：一次错误结论会被后续步骤当作前提，越往后越难纠正。做法上把验证外置为确定性工具链，模型只负责生成候选。值得借鉴的是把「可验证的断言」与「不可验证的表述」在产品里分开处理；风险是覆盖范围受限于能写出的验证器，开放式任务里可验证比例可能不高，需要如实标注不可验证部分。

reverify's pitch: stop your AI making things up — it proposes, deterministic tools decide, and every claim is checked against ground truth with evidence, with verified facts and context surviving resets^[26]. It addresses hallucination accumulating in long tasks: one wrong conclusion becomes a premise for later steps and gets harder to correct. Verification is externalised into a deterministic toolchain while the model only produces candidates. Worth borrowing is separating verifiable assertions from unverifiable statements in the product; the risk is coverage limited to verifiers you can write, so label unverifiable parts honestly.

这个项目提供一套自组织的知识库方案：完整指南、起始 vault、Agent 技能与脚本，让知识库在 Claude Code 与 Obsidian 之间自动维护^[27]。它要解决的是个人知识管理最常见的失败模式：收集很容易，整理与回顾从不发生，最后笔记库变成无人访问的垃圾场。做法上让 Agent 承担分类、链接与定期回顾，人只负责输入与判断。值得借鉴的是把「维护知识库」的重复劳动交给 Agent，并保留人类决定优先级的权力；风险是自动分类一旦混乱会持续放大错误，需要定期人工校准结构。

This project offers a self-organising knowledge base setup: a full guide, starter vault, agent skills and scripts that maintain the library across Claude Code and Obsidian^[27]. It addresses the classic failure of personal knowledge management: collecting is easy, organising and reviewing never happen, and the vault becomes an unvisited dumping ground. The agent takes on categorisation, linking and periodic review while the human supplies input and judgement. Worth borrowing is delegating the maintenance labour to an agent while keeping priority decisions human; the risk is that a bad taxonomy compounds itself, requiring periodic manual recalibration.

这个仓库是在世界模型中训练客体永久性（object permanence）的代码库^[28]，对应「物体离开视野后是否仍然存在并被正确建模」这一能力。它要解决的是视频世界模型的经典缺陷：物体一旦被遮挡或被移出画面，模型就丢失其状态，重新出现时往往已经改变。做法上通过带有遮挡与再出现的训练数据，让模型学会维持隐式的物体状态。值得借鉴的是把「视野外的状态一致性」当成独立可训练目标，而不是指望模型从预测下一帧中自然习得；风险是这类能力高度依赖训练场景的遮挡模式，迁移到真实环境的泛化仍需验证。

This repository is the codebase for training object permanence in world models^[28], the ability to keep representing an object that has left the field of view. It addresses a classic video world-model flaw: once an object is occluded or leaves frame the model loses its state, and it often returns changed. Training data with occlusion and re-entry teaches the model to maintain implicit object state. Worth borrowing is treating out-of-view state consistency as its own trainable objective rather than hoping it emerges from next-frame prediction; the risk is dependence on the occlusion patterns of the training scenes, leaving real-world generalisation unverified.

Procedura 做的是带过程控制的 Agentic 3D 建模：把文字提示变成可编辑的参数化程序，并可选地为每个部件指定材质与关节^[29]。它要解决的是生成式 3D 的可编辑性难题：多数工具产出的是不可改的网格，一旦需求变化就要重新生成，无法进入工程流程。做法上让输出是「形状即代码」的程序，改参数即可调整，并保留与 OpenUSD 等格式的衔接。值得借鉴的是把生成结果设计成可迭代的程序而不是终态资产；风险是参数化表达力有限，复杂有机形状难以用程序描述，适用范围需要界定。

Procedura offers agentic 3D modelling with procedural control: a text prompt becomes an editable parametric program, optionally with per-part materials and articulation^[29]. It tackles the editability problem of generative 3D: most tools return immutable meshes that must be regenerated when requirements change, blocking engineering workflows. Output is a shape-as-code program that responds to parameter changes and interoperates with formats such as OpenUSD. Worth borrowing is designing generated output as an iterable program rather than a terminal artefact; the risk is limited parametric expressiveness, so scope must be defined for complex organic shapes.

RoboRSI 提供一个机器人 Agent harness，带 CLI 与本地 Web 控制台，用于 LIBERO 短任务评测^[30]。它要解决的是具身 Agent 的评测工程问题：跑一次仿真评测往往需要拼装环境、脚本与日志，很难持续跑回归。做法上把执行、观测与结果展示打包成可重复运行的 harness。值得借鉴的是先建设评测 harness，再谈 Agent 改进——否则无法判断改动是否真的有效；风险是它绑定 LIBERO 这一特定基准与短任务设定，结论未必能外推到长程真机任务。

RoboRSI provides a robot-agent harness with a CLI and a local web console for LIBERO short-task evaluation^[30]. It addresses the evaluation engineering problem for embodied agents: running a simulation evaluation usually means assembling environment, scripts and logs, which makes regression runs impractical. The harness packages execution, observation and result display into a repeatable run. Worth borrowing is building the evaluation harness before improving the agent, otherwise changes cannot be shown to help; the risk is tight coupling to the LIBERO benchmark and short tasks, which may not transfer to long-horizon real-robot work.

4DCodeBench 把逆图形学（inverse graphics）变成代码生成任务：Agent 要从视频重建动态场景，产出的是可执行的图形程序，而不是一张图或一段描述^[31]。它要解决的问题是评测盲区：视觉重建做得好，不代表能理解物理——任务要求 Agent 写出物理仿真等抽象来复现形变、流体与断裂等复杂行为。数据集覆盖真实视频与合成的多样物理现象，并对前沿模型做了系统评测，结论是强的静态重建能力并不能自动迁移到动态场景的建模。对做具身与世界模型的团队，值得借鉴的是用「能否写出可运行的程序」作为理解的检验标准；限制是任务偏合成与结构化，开放式真实视频上的表现仍需观察。

4DCodeBench turns inverse graphics into a code-generation task: agents reconstruct dynamic scenes from video as executable graphics programs rather than an image or caption^[31]. It targets an evaluation blind spot: good visual reconstruction does not imply physical understanding — the task requires implementing abstractions such as physical simulation to reproduce deformation, fluid flow and fracture. The dataset spans real videos and synthetic physical phenomena, and extensive benchmarking finds that strong static reconstruction does not translate to dynamic scene modelling. For embodied and world-model teams the transferable idea is using “can it write a runnable program” as the test of understanding; the limit is a synthetic, structured emphasis.

WING 试图解决机器人数据稀缺的问题：第一人称（egocentric）人类视频里包含大量可迁移的交互经验，但直接迁移有两个障碍——从帧重建推断出的潜在动作容易被自我相机运动等无关变化主导，且人与机器人的时间动态并不一致^[32]。方法上把学习重心放在「交互」而不是「像素」：以交互为中心做谱域潜在引导，抑制视角抖动等干扰，并对齐跨实体的时间动态。对做机器人策略与视频预训练的团队，值得借鉴的是显式区分「任务相关的交互信号」与「与任务无关的观测噪声」；限制是方法依赖第一人称视频的质量与视角分布，真机部署的成功率仍需独立复现。

WING attacks robot data scarcity: egocentric human video holds abundant transferable interaction experience, but direct transfer fails for two reasons — latent actions inferred from frame reconstruction get dominated by nuisance variation such as ego-camera motion, and human and robot temporal dynamics differ^[32]. The method centres learning on interaction rather than pixels, using interaction-centric spectral latent guidance to suppress viewpoint jitter and aligning temporal dynamics across embodiments. For robot policy and video-pretraining teams the transferable idea is explicitly separating task-relevant interaction signal from task-irrelevant observation noise; the limit is dependence on egocentric video quality and viewpoint distribution.

CORNAV 面向施工现场的机器人导航，并给出了一组很不「实验室」的前提：建筑业长期缺工、生产率低下（全球每年损失超过 1.6 万亿美元），且工伤率在主要行业中居前^[33]。它指出现有语言导航系统只依赖语义场景理解，缺少施工特有的语境——建筑图纸、不断变化的施工计划与安全约束，因此无法可靠定位永久构件，也无法安全穿越在施工地。做法上把「图纸约束」与「计划感知」接入导航推理。对工作在现场级具身系统的团队，值得借鉴的是把领域约束（图纸、排程、安全规程）当成推理输入而不是后处理规则；限制是依赖 BIM/图纸的可得性与准确性，不同工地的数据条件差异很大。

CORNAV targets robot navigation on construction sites and starts from conditions far from the lab: the industry faces persistent labour shortages, low productivity costing the global economy over $1.6 trillion annually, and one of the highest injury rates^[33]. Existing language-grounded navigation relies on semantic scene understanding alone and lacks construction-specific context — architectural plans, evolving work schedules and safety constraints — so it localises permanent features unreliably and cannot navigate active jobsites safely. The approach feeds blueprint grounding and schedule awareness into navigation reasoning. For field-level embodied systems the transferable idea is treating domain constraints as reasoning inputs rather than post-hoc rules; the limit is dependence on BIM or drawing availability.

UniIntervene++ 针对在线强化学习里最实际的问题：机器人在真实环境中学习时需要人类介入，但随着能力变化，它需要的帮助类型也会变，而基于离线估计或固定规则的介入策略会变得不匹配^[34]。方法上把演化中的策略、轨迹纠正与结构化的 CodePolicy 统一建模为半马尔可夫决策过程中的 Options，并学习它们之间的关系，让一个自适应介入 Agent 决定何时自主执行、何时请求哪一种帮助。对做真实机器人 RL 的团队，值得借鉴的是把「何时求助」当成可学习的策略而不是固定阈值；限制是需要设计若干异质辅助行为作为候选，设计质量直接决定上限。

UniIntervene++ targets the most practical problem in online RL: a robot learning in the real world needs human intervention, but the kind of help it needs changes as competence evolves, so strategies based on offline estimates or fixed rules become mismatched^[34]. It formulates the evolving policy, trajectory correction and a task-structured CodePolicy as Options in a unified semi-Markov decision process and learns their relations, so an adaptive intervention agent decides when to act autonomously and which form of help to request. For real-robot RL teams the transferable idea is learning when to ask for help instead of thresholding it; the limit is that the set of heterogeneous assisted behaviours must be designed first.

这篇论文从一个反直觉的对比出发：分类器本质上就是一个策略，其期望奖励就是「给正确标签的概率」，而且因为标签已知，策略梯度是精确且平滑的——但即便如此，精确策略梯度在期望准确率上仍然输给交叉熵^[35]。作者指出原因是精确梯度是短视的：它只按「现在能换来多少」评估一次更新，而每次更新同时也决定了下一步从哪里开始，因此一次更新的价值取决于还剩多少学习空间。也就是说，优化目标不变的情况下，考虑「学习轨迹」比只看即时收益更有效。对做 RL 后训练与 Agent 优化的团队，值得借鉴的是把更新视为轨迹规划问题，引入对后续学习空间的估计；限制是论文的分析建立在分类这一可解设定上，能否推广到复杂 LLM 后训练仍需验证。

Planning to Learn starts from a counter-intuitive contrast: a classifier is a policy whose expected reward is the probability it assigns to the correct label, and because the label is known the policy gradient is exact and smooth — yet exact policy gradient still loses to cross-entropy on expected accuracy^[35]. The reason is that the exact gradient is myopic: it values an update only by what it buys now, while each update also sets where the next one starts, so an update's value depends on how much learning remains. With the objective unchanged, reasoning about the learning trajectory beats immediate gain. For RL post-training and agent optimisation teams the transferable idea is treating updates as trajectory planning with an estimate of remaining learning headroom; the limit is an analysis grounded in the tractable classification setting.

这篇论文直面机器人落地的一个硬约束：前沿模型让机器人从少量演示中学会操作成为可能，但推理延迟太高，无法用于实时控制^[36]。作者研究两条互补路线把前沿推理与低延迟本地执行接起来：一是让前沿模型自主生成演示来补充人类演示、从而训练快速本地策略，并在上下文示例中加入纠错片段（演示如何从物理错误中恢复）以提高生成可靠性；二是用稠密语言监督把执行过程与语义对齐。一个值得注意的经验是随着成功示例在上下文中累积，生成时间与成本会下降。对做机器人系统的团队，值得借鉴的是用「慢模型造数据、快模型做控制」的分工；限制是生成演示的物理正确性仍需验证，错误演示会直接污染本地策略。

This paper confronts a hard constraint in robot deployment: frontier models make manipulation from few demonstrations possible, but inference latency is too high for real-time control^[36]. Two complementary routes connect frontier reasoning to low-latency local execution: using a frontier model to autonomously generate demonstrations that supplement human ones for training a fast local policy, with corrective segments in context showing recovery from physical errors to improve reliability; and dense language supervision aligning execution with semantics. Notably, generation time and cost fall as successful examples accumulate in context. For robotics teams the transferable idea is slow models make data, fast models control; the limit is that generated demonstrations still need physical validation.

HazardWeaver 研究自然灾害分析 Agent 的「科学路线选择」问题：Agent 需要整合科学数据、模型与工具做自动化分析，但关键在于判断哪种科学方法适合当前事件、并且在可用数据与工具下真的可执行；而随着新证据与执行结果出现，这些条件会变化，Agent 必须重新考虑自己的选择^[37]。做法上把问题形式化为依赖状态的路线选择，并让 Agent 在证据更新时重新评估而不是沿用最初计划。对做科学或工程 Agent 的团队，值得借鉴的是把「方法是否可执行」当作一等判断，而不只是「方法是否正确」；限制是依赖预先整理的方法与数据元信息，新方法接入需要人工维护。

HazardWeaver studies scientific route selection for natural hazard analysis agents: the agent must integrate scientific data, models and tools, and the crux is determining which scientific method suits a given event and is actually executable with the available data and tools — and as new evidence and results arrive these conditions change, forcing the agent to reconsider^[37]. The problem is formalised as state-dependent route selection, with re-evaluation on new evidence rather than sticking to the original plan. For scientific or engineering agents the transferable idea is treating executability as a first-class judgement alongside correctness; the limit is dependence on curated method and data metadata.

NeutronGym 的设计意图很明确：用一个无法争辩的评分环境，检验语言模型 Agent 究竟是在做物理还是在背物理^[38]。它是首个面向中子仪器设计的可执行环境：Agent 通过受校验的工具搭建仪器，McStas 对搭建结果做射线追踪，评分阶梯分别考核语法、运行、结构与科学正确性，不使用 LLM 作为评委。为了控制记忆化，程序化生成族提供无限实例并保留参数区域，另有一个包含 16 个已发表仪器任务的切片，配记忆探测与沙箱。对做科学 Agent 评测的团队，值得借鉴的是用可执行仿真加分层确定性评分替代 LLM 评审，从根本上消除评委偏差；限制是环境高度专用，构建同类环境需要领域专家投入。

NeutronGym has a clear intent: use an unarguable grading environment to test whether a language-model agent is doing physics rather than recalling it^[38]. It is the first executable environment for neutron instrument design: agents build instruments through validating tools, McStas ray-traces what they build, and a level-resolved ladder grades syntax, runtime, structure and science with no LLM judge. To control memorisation, procedural families supply unlimited instances with held-out parameter regimes, plus a curated slice of 16 tasks from published instruments behind memorisation probes and a sandbox. For scientific agent evaluation the transferable idea is replacing LLM judging with an executable simulator and layered deterministic scoring; the limit is that building such environments needs domain experts.