这个仓库提供了完整的 Agent 框架：原生命令行 TUI（Terminal User Interface，终端界面）、ReAct 模式与 Agent Team 多智能体模式，一条命令在 macOS 与 Linux 上跑起来，不要求预装环境也不强依赖 Docker^[21]。它面向的是想立刻验证多智能体编排、又不想先花两天搭基础设施的开发者。实现上的关键点是「无预装」——把运行时依赖收进单一可执行包，降低了第一次跑通的门槛。值得借鉴的是它的模式切换设计：同一个 harness 下用 ReAct 处理单线程任务、用 Agent Team 处理可并行的任务；风险是框架自带约定较多，接入自有工具链时需要评估改造量。

This repo ships a full agent framework: a native terminal TUI (Terminal User Interface), a ReAct mode and an Agent Team multi-agent mode, runnable with one command on macOS and Linux, with no preinstall and no hard Docker dependency^[21]. It targets developers who want to validate multi-agent orchestration immediately rather than spend two days on infrastructure. The key implementation idea is the zero-preinstall packaging that collapses runtime dependencies into one artefact. Worth borrowing is the mode switch — one harness for single-threaded ReAct work and one for parallelisable team work; the risk is heavy built-in convention, so assess the effort to graft on your own toolchain.

这个项目通过跨平台的本地 MCP（Model Context Protocol，模型上下文协议）能力，为 ChatGPT 加上 Chrome 集成、Goal 目标、Compact 压缩与 Resume 恢复，以及持久化的多 Agent 工作流^[22]。它解决的是网页端助手的结构性短板：会话一旦超长就丢上下文，长任务中断后无法接续。实现上把状态与能力放在本地进程里，通过 MCP 暴露给模型，从而绕开纯网页会话的长度与状态限制。值得借鉴的是「把记忆与恢复放到模型外部、由协议层承载」这一架构；风险是本地服务带来新的攻击面与权限管理负担，需要自行评估安全边界。

This project uses cross-platform local MCP (Model Context Protocol) capabilities to add Chrome integration, Goal, Compact and Resume, and durable multi-agent workflows to ChatGPT^[22]. It addresses structural weaknesses of a web assistant: context loss in long sessions and the inability to resume interrupted long tasks. The design keeps state and capability in a local process exposed over MCP, sidestepping web-session length and state limits. The idea worth borrowing is carrying memory and recovery outside the model at the protocol layer; the risk is that a local service adds attack surface and permission overhead, so define the security boundary yourself.

dscode 是一个 DeepSeek 编码 Agent harness，卖点是常驻 shell、Ultra 子智能体、自动批准、Chrome MCP 与会话遥测（session telemetry）^[23]。它面向的是需要长时间在真实仓库里跑任务的工程师：常驻 shell 保住了环境状态，子智能体负责并行分支，遥测则让「这次任务为什么会失败」可被回溯。实现上的关键在于把会话级遥测当成一等公民，而不是事后加日志。值得借鉴的是「自动批准 + 遥测」的组合——放宽权限的同时留下审计轨迹；风险是自动批准在不可信仓库里有破坏性，需配合沙箱使用。

dscode is a DeepSeek coding-agent harness whose pitch is a persistent shell, Ultra subagents, auto approval, Chrome MCP and session telemetry^[23]. It targets engineers running long tasks in real repositories: the persistent shell preserves environment state, subagents handle parallel branches, and telemetry makes “why did this run fail” traceable. The key implementation choice is treating session telemetry as first-class rather than bolting on logs later. Worth borrowing is the auto-approval plus telemetry pairing, which relaxes permissions while leaving an audit trail; the risk is that auto approval is destructive in untrusted repos, so pair it with a sandbox.

rome 把自己定义为面向递归 Agent 的复利式 Agent OS（操作系统层），同时是 Grok Bot 与 Meta Muse 的开源自托管替代^[24]。它要解决的问题是消费级 Agent 的锁定：能力、记忆与工作流都存在厂商侧，用户无法迁移也无法审计。做法上把个人 AI 的记忆、工作流自动化与自托管放在同一个本地操作系统层里。值得借鉴的是「复利」这个设计目标——每次运行都应留下可复用的资产（技能、记忆、流程），而不是一次性会话；风险是自托管意味着更新、备份与安全全部自负，运维成本被转移给了用户。

rome defines itself as a compounding agent OS for recursive agents, and an open, self-hosted alternative to Grok Bot and Meta's Muse^[24]. It targets lock-in in consumer agents, where capability, memory and workflows live vendor-side and cannot be migrated or audited. The approach puts personal-AI memory, workflow automation and self-hosting into one local OS layer. Worth borrowing is the “compounding” design goal — every run should leave reusable assets such as skills, memories and flows rather than a one-off session; the risk is that self-hosting means owning updates, backups and security yourself.

EasyWAM 提供训练、微调与评测 World Action Model（世界—动作模型）的统一框架，并内置基准与 LoRA（Low-Rank Adaptation，低秩适配）微调支持^[25]。它要解决的是这个方向缺少公共基线：各家论文的动作表示、预测目标与评测协议都不一致，结果无法横向比较。做法上把训练代码、微调路径与基准收在同一套代码里，让新方法有统一的对照。对做具身模型的团队，值得借鉴的是「先统一评测再比方法」的思路；风险是框架的抽象一旦与自有模型结构不匹配，改造成本可能高于收益，接入前应先跑通基准。

EasyWAM offers a unified framework for training, fine-tuning and evaluating World Action Models, with built-in benchmarks and LoRA (Low-Rank Adaptation) support^[25]. It addresses the missing shared baseline in this area: papers use different action representations, prediction targets and evaluation protocols, so results cannot be compared. Packing training code, fine-tuning paths and benchmarks into one codebase gives new methods a common reference. For embodied teams the transferable idea is to unify evaluation before comparing methods; the risk is that the framework's abstractions may not fit a bespoke architecture, costing more to adapt than it saves.

cadence 把自己定位成一个从实时经验中学习的大脑：默认走 System 1（感知、记忆、可塑性、行动），可选开启 System 2 做自我观察与慢速纠错^[26]。它要解决的是在线学习难题：主流 Agent 靠离线训练加推理，现场反馈很少真正改变行为。做法上使用平衡网络（equilibrium networks）与局部学习规则，让模型在部署中持续更新而不必重训全量参数。值得借鉴的是快慢双系统的工程拆分；风险是这类研究型代码的稳定性与可复现性通常有限，直接用于生产需要先补评测与回滚机制。

cadence positions itself as a brain that learns from live experience: System 1 by default (perception, memory, plasticity, action) with an optional System 2 for self-observation and slow corrections^[26]. It attacks online learning: mainstream agents train offline then infer, so in-the-field feedback rarely changes behaviour. It uses equilibrium networks and local learning rules so the model keeps updating in deployment without retraining all parameters. Worth borrowing is the engineering split between fast and slow systems; the risk is that research-grade code often lacks stability and reproducibility, so add evaluation and rollback before production use.

clodfarm 的做法是种下一个任务，让一群 Claude Code Agent 自行拆分子智能体、领取工作，并按每个账号真实的 5 小时与每周用量节流^[27]。它解决的问题很具体：多 Agent 并行跑编码任务时，真正的瓶颈往往不是算力而是账号级速率限制，撞限后整批任务一起挂掉。实现上的关键点是把配额当成一等调度输入，并支持从 Claude 应用远程下达指令。值得借鉴的是把「限流感知」写进编排器而不是靠重试兜底；风险是它强绑定单一厂商的配额规则，规则一变就需要跟着改。

clodfarm's idea is to plant a mission, let a farm of Claude Code agents split it into subagents and pick up work, pacing itself against each account's real 5-hour and weekly usage^[27]. The problem is concrete: when coding agents run in parallel the real bottleneck is often account-level rate limits, not compute, and hitting one takes down the whole batch. The key design point is treating quota as a first-class scheduling input, with remote steering from the Claude app. Worth borrowing is rate-limit awareness inside the orchestrator rather than retry-and-hope; the risk is tight coupling to one vendor's quota rules.

pi-harness 声称自己是 Pi Coding Agent 的「运维超集」：完整继承 Pi 的能力，再补上可观测性、治理、故障恢复、评估与多 Agent 编排^[28]。它针对的是 Agent 从 demo 走向日常使用时的空白：能跑通不等于能运营，缺的是日志、权限、回归评测与崩溃恢复。做法上把这些能力做成围绕 Pi 的外层控制面，而不是改模型本身。值得借鉴的是这个切分——能力归 harness，运营归控制面；风险是强依赖 Pi 生态，如果团队不用 Pi，这套抽象只能当设计参考。

pi-harness claims to be an operational superset of the Pi Coding Agent: everything Pi does, plus observability, governance, recovery, evaluation and multi-agent orchestration^[28]. It fills the gap between a working demo and daily operation: running once is not the same as running reliably, and what is missing is logs, permissions, regression evals and crash recovery. The approach builds these as an outer control plane around Pi rather than changing the model. Worth borrowing is the split -- capability in the harness, operations in the control plane; the risk is tight Pi coupling, so outside that ecosystem it is a design reference only.

pi-herdsman 专注于并行编码场景的异步子智能体与集群编排：嵌套委派、后台作业与监督机制^[29]。它解决的是「一个 Agent 顶不住大任务」的组织问题：把任务拆给多个后台子智能体同时推进，再由监督层处理失败与串行依赖。实现上把子智能体做成异步、可嵌套、可监督的对象，而不是一次性的同步调用。值得借鉴的是把嵌套委派与后台执行当成编排原语；风险是并行度上去之后调试难度与成本同步上升，需要配额与终止条件配合使用。

pi-herdsman focuses on asynchronous subagents and fleet orchestration for parallel coding: nested delegation, background work and supervision^[29]. It addresses the organisational limit of one agent on a large task by splitting work across background subagents while a supervisor handles failures and sequential dependencies. Subagents are modelled as asynchronous, nestable and supervisable objects rather than one-shot synchronous calls. Worth borrowing is treating nested delegation and background execution as orchestration primitives; the risk is that more parallelism means harder debugging and higher cost, so quotas and termination conditions are mandatory.

daidocs 提出一套开放纯文本格式用于 AI 记忆：助手的长期记忆以 .dai 文件存在本地磁盘上，Claude、GPT、Gemini、Cursor 与本地模型都能读，grep 也能查^[30]。它要解决的是记忆被厂商锁死的问题——记忆换不了引擎、也审计不了。作者给出 MCP 服务与钩子接入，并声称在 LongMemEval-S 上达到 83%（GPT-4o）与 92%（Claude Fable 5），token 消耗降低约 10 倍。值得借鉴的是用最笨的格式换取可移植与可审计；风险是这些分数来自项目自述，需独立复现，且纯文本方案在超大规模记忆下检索效率存疑。

daidocs proposes an open plain-text format for AI memory: an assistant's long-term memory lives as .dai files on your disk, readable by Claude, GPT, Gemini, Cursor and local models, and greppable^[30]. It attacks memory lock-in, where memory cannot move between engines or be audited. The project ships an MCP server and hooks, and claims 83% on LongMemEval-S with GPT-4o and 92% with Claude Fable 5 at roughly 10x fewer tokens. Worth borrowing is trading an unglamorous format for portability and auditability; the risk is that those numbers are self-reported and need independent reproduction, and plain-text retrieval efficiency at very large memory sizes is questionable.

这篇论文针对 Agent 记忆的一个隐蔽缺陷：现有系统用「某条记忆对任务表现的因果效应」来决定保留哪些记忆，但估计完全依赖该记忆被检索到——从没被检索过的记忆，做任何存储层干预结果都一样，其效用根本无法识别^[31]，作者称之为检索层的 positivity violation（正性违背）。方法是把干预从存储层搬到检索层：预留固定数量的上下文槽位，按已知概率抽样记忆，从而在随机化实验的意义上估计效用，再据此学习保留策略。贡献在于把「记忆该不该留」从相关性问题纠正成可识别的因果问题；对做记忆系统的团队，直接可借鉴的是在检索阶段埋入可控随机性并记录倾向分数（propensity score），否则离线评估会系统性高估记忆价值。

This paper targets a subtle defect in agent memory: systems decide what to retain from a memory's causal effect on task performance, but that estimate depends entirely on the memory being retrieved — a memory never retrieved yields identical outcomes under any store-level intervention, so its utility is unidentifiable^[31], which the authors call a retrieval-level positivity violation. The fix moves the intervention from storage to retrieval: reserve a fixed number of context slots for memories sampled with known propensities, making utility estimable in the randomised-experiment sense and the retention policy learnable from it. The contribution is reframing retention as identifiable causality rather than correlation; for memory teams the directly borrowable practice is injecting controlled randomness at retrieval and logging propensity scores, otherwise offline evaluation systematically overstates memory value.

OneStreamer 处理流式视频交互的核心矛盾：证据必须在其相关性被知晓之前就记录下来，而同时又不能拖累实时感知^[32]。它针对的问题是流式视频 LLM 的两难——边看边记会挤占算力，先看后记又会丢掉没来得及保存的证据。方法上用共享的主动生成过程，把「与查询无关的证据记录」和「任务响应」联合训练；其主动式分层字幕记忆（Proactive Hierarchical Caption Memory）同时产出时间锚定的局部细节描述与更高层摘要，使记忆可在之后被检索复用。值得借鉴的是把「记录」与「回答」解耦但共享表示；限制是分层记忆的构建成本与长视频下的存储增长，论文未充分讨论。

OneStreamer addresses the core tension in streaming video interaction: evidence must be recorded before its relevance is known, without compromising real-time perception^[32]. The problem is the dilemma facing streaming video LLMs — recording while watching competes for compute, but recording afterwards loses evidence that was never captured. The method jointly trains query-independent evidence recording and task response through a shared proactive generation process; its Proactive Hierarchical Caption Memory emits both time-grounded local detail and higher-level summaries so memory can be retrieved later. Worth borrowing is decoupling recording from answering while sharing representations; the limit is the cost of building hierarchical memory and storage growth on long video, which the paper does not fully address.

World Observer 抓住了视频世界模型的一个结构性缺陷：模型是「以行动者为中心」的，物体一旦离开视野就没有直接证据，重新进入时状态与动力学往往已经丢失^[33]。它要解决的是持久性问题——世界模型应该持续知道视野之外发生了什么。方法上把「观察」与「行动」解耦：在生成以智能体为中心的视角同时，联合生成一个观察者视角，让模型始终保有对场景其余部分的建模。对做世界模型与具身预测的团队，值得借鉴的是在训练目标里显式加入视野外一致性，而不是只优化当前视角的下一帧；限制是双视角联合生成带来的算力开销，以及观测者视角的训练数据如何获得。

World Observer seizes on a structural flaw in video world models: they are actor-centric, so once an object leaves view there is no direct evidence of its evolution and its state and dynamics are often lost on re-entry^[33]. The target is persistence — a world model should keep tracking what happens outside the current view. The method decouples observing from acting by jointly generating a perspective actor for the agent-centric view together with an observer view, so the model always models the rest of the scene. For world-model and embodied-prediction teams the transferable idea is adding explicit out-of-view consistency to the training objective instead of only predicting the next frame from the current view; the limits are the compute cost of joint dual-view generation and where observer-view training data comes from.

这篇论文研究从观察中推断机器人伙伴的物理约束，再据此完成零样本（zero-shot）协同：当一台机器因硬件退化或执行器故障而无法完成某些动作、而伙伴并不知情时，协同会失败^[34]。难点在于演示只显示受限机器人「做了什么」，而不是它「本来能做什么」——可从已知约束到行为的映射是多对一的。做法是让辅助机器人观察目标机器人与第三台机器协同的过程，反推出能力边界，再迁移到新任务上协同。对做多机器人或人机协同的团队，值得借鉴的是把「推断伙伴能力」当成显式的感知任务而不是靠通信假设；限制是它依赖可观察的协同演示，缺少演示时如何退化尚不清楚。

This paper studies inferring a robot partner's physical constraints from observation to achieve zero-shot coordination: when a robot cannot reliably perform certain actions because of hardware degradation or actuator faults — and its partner does not know — coordination breaks^[34]. The difficulty is that a demonstration shows what the constrained robot did, not what it could have done, because the constraint-to-behaviour mapping is many-to-one. The method has a helper infer the capability boundary from watching the constrained robot coordinate with a third robot, then transfer that inference to a new task. For multi-robot and human-robot teams the transferable idea is treating partner-capability inference as an explicit perception task rather than assuming communication; the limit is its dependence on observable demonstrations.

LLM2Jev 回答一个很实际的问题：通用 LLM 本身能不能直接当决策模型用，什么时候才真的需要微调^[35]。所谓 Jev 式决策模型，是指直接返回预定义选项上的类别概率分布、不生成自由文本，让上游软件可以直接消费输出——这正是路由、分类、门控这类 Agent 判断层需要的接口。方法上保持原有架构，从下一个 token 在带括号编号上的概率里抽取校准后的决策，既给出免训练推理方法，也给出一个用树分解 listwise 损失优化候选选择的微调目标。对做 Agent 判断层的团队，价值在于先用免训练路径测出基线，再决定是否值得为微调付成本；限制是校准质量与选项集合设计强相关，选项改动后需要重新评估。

LLM2Jev answers a practical question: can a general LLM already serve as a decision model, and when is fine-tuning actually necessary^[35]. A Jev-style decision model returns categorical probability distributions over predefined options without generating free-form text, so upstream software can consume the output directly — exactly the interface routing, classification and gating layers in an agent need. The method preserves the architecture and extracts calibrated decisions from next-token probabilities over bracketed numeric identifiers, offering both a training-free inference recipe and a fine-tuning objective that optimises candidate selection with a tree-factorised listwise loss. For judgement-layer teams the value is measuring a training-free baseline before paying for fine-tuning; the limit is that calibration quality depends heavily on option-set design, which must be re-evaluated whenever options change.

AgSpec 针对编码 Agent 流水线的推理效率：检索式投机解码（speculative decoding，用已有文本的续写来起草 token）天生适配编码 Agent，因为代码、日志与之前的尝试会被反复复现^[36]。但现有方法在 Agent 场景下失效，原因是可复用的文本很多不在语料里、或存储形式与 Agent 实际输出不一致，且它们假设的起草长度忽略了「接受长度随 Agent 变化、并随轮次漂移」这一点。AgSpec 的贡献在于提供对应语料并让起草长度自适应。对跑大规模编码 Agent 的团队，值得借鉴的是把重复内容当成可缓存的推理加速资产；限制是收益依赖任务重复度，探索型任务的加速幅度会明显下降。

AgSpec targets inference efficiency in coding-agent pipelines: retrieval-based speculative decoding (drafting tokens by copying continuations from existing text) suits coding agents because code, logs and earlier attempts get reproduced repeatedly^[36]. Existing methods fall short in agent settings because much reusable text is missing from their corpora or stored in a form that differs from what the agent emits, and their assumed draft lengths ignore that accept length varies across agents and drifts over turns. AgSpec supplies the matching corpus and adapts draft length accordingly. For teams running coding agents at scale the transferable idea is treating repeated content as a cacheable inference asset; the limit is that gains depend on task repetitiveness and shrink on exploratory work.

SkeleWAM 对 World Action Model（世界—动作模型）的表征方式提出质疑：现有模型预测视频或学到的视觉隐变量，交互几何只是被隐式编码，还夹带了与控制无关的外观信息^[37]。它要解决的是动作学习被外观噪声稀释的问题。方法是把操作场景表示成稀疏 3D 骨架——机器人关节、物体中心与交互点——由当前 RGB-D 观测与机器人本体感知在线构建，成为动作生成与未来骨架预测共享的几何状态；未来骨架预测反过来为动作学习提供额外的几何监督。对做操作策略的团队，值得借鉴的是用显式几何状态替代像素级预测目标以降低表征负担；限制是骨架抽象会丢失外观相关但控制相关的线索（如材质、柔性变形），需要按任务评估。

SkeleWAM challenges how World Action Models represent scenes: existing models predict video or learned visual latents, encoding interaction geometry only implicitly and retaining appearance information unrelated to control^[37]. The problem is that action learning gets diluted by appearance noise. The method represents a manipulation scene as a sparse 3D skeleton — robot joints, object centres and interaction points — built online from current RGB-D observations and robot proprioception, giving a single geometric state shared by action generation and future skeleton prediction; predicting the future skeleton then supplies extra geometric supervision for action learning. For manipulation teams the transferable idea is replacing pixel-level prediction targets with an explicit geometric state to cut representational load; the limit is that skeleton abstraction can drop appearance-linked but control-relevant cues such as material and soft deformation.

这篇论文给出了一个反直觉但可复现的结论：预训练 Transformer 只用很少的深度去完成上下文的引用跟随，13 个基础模型可靠跟随的行数只有 1.4–3.6 行，单纯堆叠预训练的循环层几乎没用^[38]。它针对的是长上下文推理里「模型看起来读完了、其实没有传递」的问题。方法是冻结全部模型权重，只在某个早期层训练一个 rank-8 的 LoRA（Low-Rank Adaptation，低秩适配），把计算接力下去：Qwen3-8B 在 24 行链式引用上的精确率从 15.5% 提升到 99%，训练更久的 LoRA 可到 50 行；Ouro-1.4B 四次循环后到 60 行、八次循环后至少 160 行。对做长上下文 Agent 的团队，参考价值是用极小的参数预算去修复推理深度，而不是一味加长窗口；限制是任务形态是链式引用，能否迁移到开放式长文推理仍待验证。

This paper reaches a counter-intuitive but reproducible conclusion: pretrained transformers use very little of their depth to follow references in context, with thirteen base models reliably following only 1.4–3.6 lines, and extra pretrained loops adding little^[38]. It targets the long-context failure where a model appears to have read the input but has not propagated it. The method freezes all weights and trains a rank-8 LoRA (Low-Rank Adaptation) at one early layer to keep the computation going: Qwen3-8B moves from 15.5% to 99% exact accuracy on 24-line chains, a longer-trained LoRA reaches 50 lines, and Ouro-1.4B reaches 60 lines after four loops and at least 160 after eight. For long-context agent teams the reference is fixing reasoning depth with a tiny parameter budget instead of only extending the window; the limit is that the task is chained reference following, leaving transfer to open-ended long-form reasoning unverified.