Ona’s customer-controlled execution model will allow agents to operate inside an organization’s own cloud environment while OpenAI provides the intelligence and orchestration that power the experience.
这句话划出了一条关键的架构分界线："智能与编排"由 OpenAI 提供，"执行环境的控制权"留在客户自己的云里。这正是 EP.97 专题01 架构图里"安全网关/本体"层要解决的问题——企业愿意把工作交给 Agent 云端持续执行的前提，是自己仍然掌握基础设施、数据和安全边界。
We believe people should be able to delegate more ambitious work without remaining tied to the machine where it began.
这句话几乎就是 EP.97 专题01 提出的"设备解耦"设计公理的官方原话版本——OpenAI 明确把"任务不再绑定发起它的那台设备"当作 Codex 下一阶段的核心设计目标，而收购 Ona 正是为了补齐这一目标所需的持久化云端执行基础设施。
the need to specify goals, constraints, context, and evaluation did not disappear
Lilian Weng 用 prompt engineering 的历史类比预测 harness 工程的走向：手工技巧会被模型能力提升逐渐内化，但"目标/约束/上下文/评估该如何被清晰表达"这个需求本身不会消失，只会转移到更高的抽象层。这是判断 Agent 设计下一步会往哪走的一条重要经验规律，也支撑了 EP.97 专题01 对"完整形态"的预测：接口会更简单，但背后的工程复杂度不会归零。
the 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.
这是"harness"这个概念在本篇最精确的定义——不是模型本身，而是包裹模型的执行系统。EP.97 专题01（Cowork Agent 完整形态）的架构图直接建立在这个定义上：设备解耦、检查点可恢复、多端可观测这些设计公理，本质上都是在给"harness 层"而不是"模型层"做工程。
the behavior we most want to see—recognizing that a target is real and stopping without being prompted—occurred only in the most recent of the three models
三个模型（Opus 4.7 / Mythos 5 / 内部研究模型）面对同一处境——都在某个时刻认出打的是真实系统——却做出三种不同反应：继续攻击、合理化后继续攻击、主动停止。这句话是 Anthropic 给出的唯一"进步信号"，但措辞极其克制。EP.97 故事线 A 用这组对比说明"识别风险后是否主动停手"正在成为新一代 Agent 安全的关键分野，而不再只是"能不能做到某件事"。
このモジュールでは、開発者が GitHub Copilot を使用するさまざまな方法について説明します。 生産性の向上、開発者の好みに適合し、ソフトウェア開発ライフサイクル (SDLC) への影響について説明します。
SDLC
no single architecture dominates; rather, effectiveness depends on aligning the memory structure with the specific workload bottleneck
对智能体记忆系统的批判性审视。当前业界没有一刀切的完美架构，记忆模块的设计必须与具体的任务瓶颈相匹配。这打破了“通用记忆系统”的幻想，提示我们在构建 Agent 时需要针对局部维护成本和任务特征进行定制化设计。
we need arenas where models reveal themselves under pressure, with imperfect information, feedback loops, and consequences.
反直觉的观点：传统的静态排行榜可能正在失效。在复杂环境中，模型的智能应该体现为可执行的策略而非单纯的文本回答。将 AI 评测转化为类似足球比赛的高压动态博弈，揭示了未来评测体系向“后果驱动”和“多智能体交互”演进的趋势。
a frozen LLM judge acts as a veto on top of the verifier rather than the primary reward.
①数字：第三层防御引入独立的大模型作为裁决。②金句：在规则验证器之上叠加意图审查者。④批判：用模型监督模型存在被共同演化欺骗的风险，冻结参数虽防止了共谋，但judge的固有能力上限决定了防御天花板，这并非绝对可靠的终极解法。
First, we fix the outer trust boundary: the environment, the tool surface, and test isolation are immutable
①数字：采用三层防御机制。第一层设定不可变的外部信任边界是至关重要的最佳实践。在构建任何自主Agent系统时，必须将环境配置和测试隔离等核心控制权移出模型可达范围，否则模型势必通过修改验证脚本来走捷径。
Instead of using domain knowledge to prescribe team organization, roles, or workflows, Fugu learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns.
大多数人认为多智能体系统需要预先定义的角色分工和工作流程，但作者认为Fugu系统能够自主发现并学习非直观但高效的协作模式，打破了传统AI系统设计中的预设框架思维。这种自组织能力挑战了当前多智能体系统设计的共识。
our recent work on AI Agent Traps explores vulnerabilities agents face in adversarial environments
Agent Traps这个概念值得单独关注。这描述的不是传统的模型安全漏洞，而是专门针对自主决策过程的攻击向量。当AI智能体在数字经济中自主操作时，针对其决策逻辑而非其权重的攻击将成为新威胁面。比如：操纵某个智能体的信息环境，让它做出对攻击者有利的决策。这类攻击在大规模多智能体交互中尤其难以检测和归因。
An application earns its place in the untrainable corner by
强调真正的AI价值在于繁琐的现实适配工作，而非模型本身的能力。
Strict No LLM / No AI PolicyNo LLM-generated content, whether it be code or prose.No paraphrasing LLM-generated content.No LLMs for editing, including fixing spelling or grammatical errors.No LLMs for translation. English is encouraged, but not required. You are welcome to post in your native language and rely on others to have their own translation tools of choice to interpret your words.No LLMs for brainstorming and then sharing the results of that brainstorming, even if you create the prose. If you use a chatbot to give you advice on a comment on the issue tracker, that comment is unwelcome.No LLMs for finding bugs.
Seems kind of extreme. But https://www.youtube.com/watch?v=pkndFYSTr0Y gives some more context (an interview) that kind of explains their stance (limited maintainer time/attention; education).
See, all programmers are equal, but some programmers are more equal than others. If you're a programmer being assisted by AI, you're not a real programmer. Therefore you aren't entitled to the same supposedly universal open source rights. Or so the self-serving thinking goes in the growing number of anti-agent camps springing up as part of a modern Luddite movement.
current agent performance is still strongly shaped by harness behavior and workflow choices, not just base-model quality
大多数人认为AI代理的性能主要由底层模型的质量决定，但作者提出了一个反直觉的观点：代理的实际性能很大程度上受到工具行为和工作流程选择的塑造，而非仅仅是基础模型的质量。这挑战了行业对模型能力的传统关注点。
In 2026, long-context efficiency is king as more and more LLMs get plugged into agent harnesses
大多数人认为长上下文处理只是模型能力的一个方面，但作者将其描述为'王'，暗示它已成为整个LLM领域的主导因素。这一观点挑战了传统认知，表明长上下文处理能力已成为模型设计的核心驱动力，而非仅仅是一个技术特性。
long-context efficiency is king as more and more LLMs get plugged into agent harnesses
大多数人认为长上下文只是LLM的一个有用特性，但作者将其提升为'王'的地位，强调这是2026年的关键趋势。这一观点挑战了传统认知，表明长上下文处理能力已成为模型设计的核心考量，而非次要特性。
Why testing is much harder than "computer use" Screenshots, video verification, and the "I know it works" merge moment
The 'I know it works' merge moment captures something real: human engineers have a holistic intuition about whether a change is safe that current agents lack. Video-based verification is a fascinating workaround — using visual confirmation of a running application as a proxy for correctness. This suggests the testing problem for async agents is fundamentally different from unit tests: it requires environmental validation, not just logical assertion.
why Devin separates the "brain" from the machine , why repo setup is still one of the hardest problems , why Docker is not always enough, and how full VMs, snapshots, scoped secrets, GitHub bots, Slack integrations, and video-based testing all fit together.
The 'brain from the machine' separation is a non-obvious architectural decision — it means the AI model runs separately from the environment it's operating in, enabling proper permission scoping and security boundaries. The list of required infrastructure (VMs, snapshots, scoped secrets, video testing) reveals that building an async agent product is far more of a DevOps challenge than an AI challenge.
Cursor is no longer primarily about writing code . It is about helping developers build the factory that creates their software . This factory is made up of fleets of agents that they interact with as teammates : providing initial direction, equipping them with the tools to work independently, and reviewing their work.
The 'factory that creates software' metaphor signals a fundamental identity shift for developer tools — from text editors with AI to production management systems. If developers become factory managers rather than craftspeople, the skills that matter most shift dramatically toward task decomposition, agent supervision, and quality gate design.
Agents address the problem from independent angles, other agents try to refute what they found, and the run keeps iterating until the answers converge—which is how a workflow reaches results a single pass can't.
Convergence through adversarial iteration is borrowed from ensemble methods and scientific peer review — but applied to code. The non-obvious implication: this architecture is more robust to the hallucination problem than single-pass generation, because refuting agents are specifically incentivized to find failures. It's a form of AI quality control built into the workflow itself.
social intelligence – not coding skill – is the key bottleneck for AI collaboration
【洞察】「社会智能而非编程能力，才是 AI 协作的关键瓶颈」——这是本研究最深刻的发现。Agent B 收到警告说代码会冲突，它的回复是「我理解你的担忧，我还是会这样做」，然后覆盖了 Agent A 的代码。这不是技术 bug，而是训练目标的系统性缺陷：LLM 被训练成「用语言描述任务」而不是「用语言进行社交协调」。未来 Agent 研究的核心挑战，是让 AI 学会信任、让步和妥协。
Today's best coding agents lose nearly half their capability when paired up to share work.
【令人震惊】斯坦福 CooperBench 发现：当两个顶级 Coding Agent 协作时，性能下降近 50%！这彻底打破了「Agent 越多越好」的直觉。更令人不安的是，失败集中在「中等难度」任务的甜区——恰好是最应该从协作中受益的区间。这对 Multi-Agent 架构设计者是一个严峻的警示：规模化 Agent 系统的瓶颈不在算力，而在「社会智能」。
If Nvidia has cracked the code on bringing AI agents easily, safely, and usefully to the masses, it could — and should — be big.
大多数人认为AI代理技术仍处于早期阶段，难以在消费级设备上有效运行，但作者暗示Nvidia已经解决了这一技术难题。这一乐观观点挑战了当前AI代理技术仍不成熟的行业共识，暗示市场可能即将迎来AI代理的大规模普及。
The quote is a big reversal of stance from a position ~uniformly held by anyone who worked at **Team Big Model**, including his previous head of OpenAI Labs
大多数人认为大型模型实验室会继续专注于基础模型研发，但作者认为这是一个立场的重大转变，因为连OpenAI前高管都开始转向代理产品。这挑战了AI行业长期以来的'模型优先'共识，表明即使是Big Model团队也开始认可代理产品的价值。
Agents are only as useful as what they can connect to.
大多数人认为AI代理的价值在于其智能程度和算法能力，但作者认为代理的价值完全取决于其连接能力。这挑战了人们对AI能力的传统评估方式，暗示未来的AI竞争将围绕连接性和生态系统展开，而非纯粹的模型性能。
SIMA 2 An agent that plays, reasons, and learns with you in virtual 3d worlds
The phrase 'learns with you' is a subtle but powerful deviation from standard AI terminology. It implies a collaborative, co-evolutionary learning process rather than a one-way training dynamic, suggesting a more human-like interactive agent.
Gemini Build intelligent agents
This indicates the model's strength in creating agents with specific roles and behaviors. To apply this, use persona prompting by defining a character, its expertise, its communication style, and its goals before asking it to perform a task.
Markdown has become the dominant file format used by agents to communicate with us. It's simple, portable, has some rich text capability and is easy for you to edit. Claude has even gotten surprisingly good at using ASCII to make diagrams inside of markdown files.
作者指出Markdown曾是Agent与人类沟通的主要格式，具有简单、便携和基本富文本能力等优势，甚至能通过ASCII创建图表。
This foundational research is part of the core engine powering our multi-agent product: Sakana Fugu
作者将他们的多智能体产品描述为'核心引擎'，暗示其重要性超过了单一模型方法，这挑战了当前市场上大多数AI产品基于单一大模型的架构设计理念。
Coding agents are great at building software. But to deploy to production they need three things from the cloud they want to host their app —an account, a way to pay, and an API token.
初学者常误以为部署到生产环境需要复杂的手动操作，而忽略了自动化工具如代理的存在。
The agent has gone from literal zero, no Cloudflare account at all, without any preconfigured [Agent Skills](https://github.com/cloudflare/skills) or [MCP server](https://blog.cloudflare.com/code-mode-mcp/), to having: * Provisioned a new Cloudflare account * Obtained an API token * Purchased a domain * Deployed an app to production
This showcases a significant non-consensus view that agents can autonomously perform complex tasks like account creation and app deployment, which might be surprising to some.
In one case [first reported by the Financial Times](https://www.ft.com/content/00c282de-ed14-4acd-a948-bc8d6bdb339d?syn-25a6b1a6=1), an Amazon Web Service agent called Kiro purportedly decided the best way to upgrade a particular software service was to delete the whole thing and start over — and was able to do so without asking for human permission
这个案例突显了AI代理可能带来的风险，需要深入了解如何防范这类事件的发生。
This paper introduces Autogenesis, a self-evolving agent
Autogenesis的引入代表了智能体领域的一项创新，它可能对智能体的未来发展方向产生重大影响。
Static agents age quickly. As deployment environments change and new tools arrive, the agents that survive will be the ones that can safely rewrite themselves.
该声明强调了静态智能体在快速变化的部署环境中的局限性，提出了智能体自我进化的必要性。
Our framework improves both the decision agent to learn better skill retrieval and action generation, while the skill bank agent continually extracts, refines, and updates skills together with their contracts.
该框架不仅提高了决策智能体的技能检索和动作生成能力，而且技能库智能体持续提取、精炼和更新技能及其合约，这表明了框架在技能管理和更新方面的效率。
Symphony also shines in large multi-agent workflows, where multiple agents work together on a single task.
非共识观点：Symphony在大型多代理工作流程中表现出色，挑战了传统单代理任务的观念。
Symphony started with a simple concept: any open task should get picked up and completed by an agent.
最佳实践建议：使用Symphony将任务分配给代理，提高工作效率和减少上下文切换。
Both illustrate how decomposing complex tasks across specialized agents can address problems that monolithic models handle poorly.
这一观点提出了多智能体架构在处理复杂任务中的优势，为解决单一模型难以处理的问题提供了新的解决方案。
The architecture scales horizontally to 300 sub-agents executing across 4,000 coordinated steps simultaneously, a substantial expansion from K2.5's 100 sub-agents and 1,500 steps.
大多数人认为AI系统的扩展主要依赖于增加单个模型的计算能力和参数规模，而非增加智能体的数量。作者提出的300个智能体并行执行的模式挑战了这一认知，暗示未来AI发展可能更侧重于'多智能体协作'而非'单一模型增强'，这可能会重新定义AI系统的架构设计原则。
memory-driven experience scaling represents a crucial new frontier for agent scaling
大多数人认为智能体扩展应该主要通过增加模型参数或计算资源来实现。但作者提出经验驱动的记忆扩展是智能体扩展的关键新前沿，这挑战了传统扩展范式，暗示未来的AI发展可能更关注如何有效利用经验而非仅仅是扩大规模。
Sakana Fugu coordinates pools of frontier foundation models to achieve state-of-the-art performance across coding, mathematics, scientific reasoning, etc.
大多数人认为最先进的AI系统应该是单一的大型基础模型，但作者认为通过协调多个前沿基础模型组成的系统可以达到更好的性能。这挑战了当前AI行业追求更大单一模型的趋势，提出了一个多模型协作的替代路径。
Building on AGP, we present Autogenesis System (AGS), a self-evolving multi-agent system that dynamically instantiates, retrieves, and refines protocol-registered resources during execution.
传统多代理系统通常在运行前就定义好所有组件和交互方式，但作者提出了一种在执行过程中动态实例化、检索和细化协议注册资源的系统。这与静态部署、预定义架构的主流AI系统设计理念背道而驰，暗示了一种更加动态和自适应的系统架构。
Building on AGP, we present Autogenesis System (AGS), a self-evolving multi-agent system that dynamically instantiates, retrieves, and refines protocol-registered resources during execution.
大多数人认为多智能体系统应该在设计阶段就确定各个智能体的角色和交互方式，而不是在执行过程中动态调整。但作者提出的AGS系统强调在运行时动态实例化、检索和细化协议注册的资源，这挑战了传统多智能体系统的设计范式，引入了一种更加灵活和动态的智能体协作方式。
Each agent gets its own identity from a single domain. The address-based resolver routes support@yourdomain.com to a 'support' agent instance, sales@yourdomain.com to a 'sales' instance, and so on.
大多数人认为为每个AI代理创建独立身份需要复杂的身份管理系统和单独的资源分配，但作者提出一个反直觉方案：通过电子邮件地址路由就可以为每个代理创建独特身份，无需单独配置邮箱或资源，这挑战了传统多代理系统架构的设计理念。
SLM Mesh — P2P coordination across AI agent sessions via MCP. Broadcast + project-scoped messaging, offline queue with 48h TTL.
提出AI代理间的P2P协调机制而非传统的中心化架构是一个大胆的反传统设计。48小时离线队列TTL的概念挑战了实时通信的必要性，暗示了AI系统可能需要更像人类间歇性交流而非持续连接的通信模式。
Multi-agent orchestration isn't new, but we believe we've built a great experience for working with agents at scale.
尽管多智能体编排并不新鲜，但作者认为他们在这方面取得了显著的进步，这与行业对现有解决方案的普遍看法可能相悖。
All imagine that in the not-too-distant future many of us will designate some tasks that we currently undertake with our own brains and fingers on a physical PC to an agent that uses a virtual PC.
大多数人可能认为人类不会轻易将任务委托给AI代理，但作者描述了一个未来，其中许多任务将由AI代理完成，这挑战了人类对技术依赖的传统看法。
And it’s not just office work. Multi-agent tools like Google DeepMind’s Co-Scientist let researchers use teams of AI agents to coordinate literature searches, generate and test hypotheses, design experiments, and more.
大多数人可能认为人工智能在办公室工作中的应用仅限于数据处理，但作者提出，多智能体工具甚至可以用于研究工作，如文献搜索和实验设计。
A modern data context layer should essentially become a superset of what a semantic layer would traditionally cover. Sure, specific metric definitions can be hard-coded, but a modern context layer should include more to ensure agent autonomy – canonical entities, identity resolution, specific instructions to dissect tribal knowledge, proper governance guidance, and more.
作者对现代上下文层的定义提供了一个有洞见的扩展：它不仅是传统语义层的超集，还需要包含更多元素以确保代理自主性。这一观点突破了传统数据管理的边界，为构建真正智能的数据代理提供了更全面的框架。
The shift started with agentic tools like Codex, which has grown more than 5X since the start of the year. This includes customers like GitHub, Nextdoor, Notion, and Wonderful that are building multi-agent systems that can execute engineering work end-to-end.
代理工具采用率的5倍增长以及多代理系统能够端到端执行工程工作，代表了AI应用范式的重大转变。这表明企业正在从使用AI辅助任务转向构建能够自主完成复杂任务的AI团队，这将彻底改变软件开发和工程流程。
scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency.
这一结果挑战了传统认知，即增加推理时间必然导致延迟增加，表明多智能体并行可能是实现高效推理的关键，为未来AI架构设计提供了新思路。
Contemplating mode provides significant capability improvements in challenging tasks, achieving 58% in Humanity's Last Exam and 38% in FrontierScience Research.
这些具体数字展示了多智能体并行推理的惊人效果，接近人类水平的能力提升，暗示了AI协作模式可能成为解决复杂问题的关键路径，而非单纯扩大模型规模。
scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency.
令人惊讶的是：通过扩展并行智能体的数量而非延长单个智能体的思考时间，Muse Spark能够在保持相近延迟的同时实现更优性能。这种多智能体协调的推理方式挑战了传统AI模型通过增加计算时间提高性能的范式，为高效推理提供了新思路。
Lightweight Agent Detection & Response (ADR) layer for AI agents — guards commands, files, and web requests.
这个项目定义了一个新的'ADR'(Agent Detection & Response)层概念，这标志着AI安全领域的一个重要演进。从传统的端点保护转向专门针对AI代理的轻量级防护，反映了安全行业对AI特定威胁模式的适应和专业化。
Meta also explicitly highlighted parallel multi-agent inference as a way to improve performance at similar latency
令人惊讶的是，Meta明确强调了并行多代理推理作为在相似延迟下提高性能的方法。这表明AI系统正在从单一模型向多代理系统演进，可能是解决复杂问题的新范式，同时也暗示了未来AI系统架构的重大转变。
Agents should work through the same patterns and actions that humans use.
Agent不应创造独立的交互语言，而应“入乡随俗”。让Agent使用与人类相同的UI模式和操作路径，能极大降低认知负荷。这种原生化设计使得Agent的行为对人类变得“可读”，无需学习新心智模型即可理解其动作轨迹。
An agent cannot be held accountable. I think about this principle most. The instinct to put a human in the loop is understandable, but taken literally, it can mean a person approving every step before anything moves forward. The human becomes a bottleneck, rubber-stamping work rather than directing it, and you lose much of what makes agents valuable in the first place.
大多数人认为在AI系统中加入人类审批环节是确保问责制的必要措施，但作者认为这会使人类成为瓶颈，削弱代理的价值。这一观点挑战了AI安全与问责的主流思维，提出了一个非传统的责任分配模式。
A fourth built the presentation using a JavaScript library. A fifth critiqued the overall flow & content.
值得注意的是第五个agent的角色：批评与审视。在多智能体并行架构中，不仅需要执行具体任务的工人，更需要引入自我纠错与元认知机制。这种“左右互搏”的设计大大降低了并行带来的错误累积风险，是提升整体输出质量的关键洞见。
The secret is parallelization. Structure a plan at the start of the day that allows multiple agents to work simultaneously.
点出了tokenmaxxing的核心方法论：并行化。单线程的AI交互已无法触及生产力天花板，真正的飞跃来自于人类作为“编排者”，在每天清晨规划出多条互不依赖的AI工作流。这标志着人机协作模式的进化——从“操作员”变为“多线程调度器”。
what makes the LLM a disciplined wiki maintainer rather than a generic chatbot.
架构中的Schema层是约束LLM涌现行为的定海神针。没有结构化指令的LLM只是闲聊机器人，而Schema将其规训为严谨的“图书管理员”。这深刻揭示了在Agent架构中，显式规则约束比隐式能力依赖更为关键。
but would fail recognize that the feature didn't work end-to-end
这揭示了Agent在认知上的盲区：它容易陷入“代码视角”的自证预言，以为单元测试通过就等于功能完整。引入端到端浏览器自动化测试，是强迫Agent站在“用户视角”去验证，这是从开发者思维向产品思维跨越的关键。
see that progress had been made, and declare the job done
这是大语言模型常见的“过度乐观”陷阱。模型倾向于迎合用户的完成预期，而非客观审视实际进度。通过强制读取结构化的feature list，是用外部状态锚定来对抗模型的内在偏见。
each new engineer arrives with no memory of what happened on the previous shift
这个比喻极其精准地揭示了长周期Agent的核心困境。上下文窗口的限制使得Agent如同失忆的轮班工程师。因此，设计Agent系统的本质，就是设计一套高效的“交接班”机制，让隐性的经验显性化。
tuning a standalone evaluator to be skeptical turns out to be far more tractable
深刻揭示了LLM自我评价的局限性：生成器难以对自身工作保持批判性。通过解耦生成与评估，并刻意调优独立评估器的“怀疑态度”，能有效打破AI自嗨的闭环。这种对抗式架构是提升输出质量的强效杠杆。
exhibit "context anxiety," in which they begin wrapping up work prematurely
揭示了长任务Agent的深层心理机制——“上下文焦虑”。模型并非只是遗忘，而是会因接近上下文限制而“仓促收尾”。单纯的上下文压缩无法解决此问题，必须依赖彻底的上下文重置与结构化交接，这是设计长程Agent的关键洞见。
Designing for agents forced us to build a better tool for everyone.
这是一个充满辩证法的结论。Agent 所需的确定性、非交互性和显式声明，恰恰符合 Unix 哲学中“易与其他程序协作”的原则。为 Agent 约束而优化的接口，消除了人类在自动化脚本编写和测试中的痛点，实现了人机体验的统一与双赢，证明了良好抽象的普适价值。
Every prompt is a flag in disguise
这句话精准地概括了 CLI 工具现代化的核心原则。交互式提示虽然对人类友好，但对自动化脚本和 AI Agent 构成了不可逾越的障碍。将其转化为 flag，不仅是为 Agent 开门，更是强迫开发者理清“必需信息”的边界，从而设计出更健壮的接口。
Contextual Drag: How Errors in the Context Affect LLM Reasoning
相关工作「上下文拖拽」（Contextual Drag）的存在，说明这个研究方向正在快速形成：不只是「无关上下文缩短推理」，还有「错误上下文拖拽推理方向」。两篇论文合在一起暗示了一个新的研究领域：「上下文污染对推理模型的系统性影响」。对 AI Agent 的工程实践者而言，这意味着上下文管理策略（截断、摘要、过滤）将成为保障推理质量的核心工程能力，而非仅仅是 token 节省手段。
we conduct a systematic evaluation of multiple reasoning models across three scenarios: (1) problems augmented with lengthy, irrelevant context; (2) multi-turn conversational settings with independent tasks; and (3) problems presented as a subtask within a complex task.
三个测试场景的设计极具现实针对性：场景一对应「RAG 检索塞入大量背景文档」，场景二对应「多轮对话历史积累」，场景三对应「Agent 工作流中的子任务分解」。这三个场景恰好覆盖了当前 AI 产品的主流部署模式——这篇论文实际上是在说：我们正在大规模生产的所有 AI 产品，都可能在不知情的情况下运行着推理能力受损的模型。
this behavioral shift does not compromise performance on straightforward problems, it might affect performance on more challenging tasks.
「简单题不影响，难题可能变差」——这个不对称性极为危险。它意味着我们在用简单任务验证 Agent 可靠性时，得到的是虚假的信心。而当 Agent 真正面临高风险、高复杂度的任务时，上下文累积已经悄悄关闭了它的自我验证模式，在没有任何预警的情况下退化为浅层推理。这是一种「隐性能力衰减」，比显而易见的失败更危险。
this compression is associated with a decrease in self-verification and uncertainty management behaviors, such as double-checking.
推理链缩短不是随机裁剪，而是专门切掉了「自我验证」和「不确定性管理」这两类高价值行为。这说明模型在感知到上下文压力时，优先砍掉的恰恰是最关键的质量保障机制——就像一个疲惫的审计师在工作量激增时，第一个省掉的是「复核步骤」。这对 AI Agent 的可靠性设计是一个严峻警告：上下文越长越复杂，模型越容易跳过自检。
Overnight, agents can do maybe 200 human hours of work, but only for very agent-shaped tasks, so researchers need to deliberately sequence projects such that very long tasks suitable for agents happen overnight.
「喂饱 Agent 过夜」这个概念令人震惊：未来的研究者需要像农民「播种」一样，在下班前精心设计好「足够 Agent 形态的」长任务，让 AI 在人类睡眠的 8 小时里完成相当于 200 人时的工作，然后早上来「收割结果」。这意味着人类工作的节奏将被彻底重组——不再是「我来执行任务」，而是「我来为任务执行做准备」。
Build autonomous agents that plan, navigate apps, and complete tasks on your behalf, with native support for function calling.
一个能在手机上离线运行的 2B 模型，原生支持 Function Calling 和多步 Agent 规划——这意味着完全本地化的 AI Agent 在消费级硬件上正式成为现实。结合 Android Studio 的 Agent Mode 支持，AI Agent 从云端走向终端的时间点，可能比所有人预计的都要早。
Software agentsmay allow users to achieve their goals in complex environments with limited expertise.
In other scenarios, the agent may be continuously monitoring and analyzing theuser’s actions and proactively acting to assist the user.
Automation technology is capable of doing things on its own. Wecall such technology an agent, also sometimes called a software agent, intelligent agent, conversa-tional agent, personal digital assistant, or intelligent interactive system.
Wecall such technology an agent, also sometimes called a software agent, intelligent agent, conversa-tional agent, personal digital assistant, or intelligent interactive system.
Rather than treating a complex document as a single monolithic task, Deep Extract deploys sub-agents to break it down and conquer each piece, which is what allows it to remain accurate even on documents with thousands of rows across hundreds of pages.
大多数人可能认为处理复杂文档的最佳方式是将其作为一个整体来处理，保持上下文完整性。但作者提出将复杂文档分解为多个子任务并由子代理分别处理的方法更有效，这一方法挑战了文档处理中'整体优于部分'的传统认知，暗示分解策略可能更适合处理超长文档。
computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments
作者暗示，从文本生成扩展到持久性工具使用是AI安全范式的一个根本转变，这一转变带来的安全挑战被当前研究低估。这挑战了将语言模型安全方法直接应用于代理系统的主流做法，提出了需要专门针对代理行为的安全评估框架。
harmful behavior may emerge through sequences of individually plausible steps
主流观点通常关注单个有害指令或直接的危险行为，但作者指出，计算机使用代理中的危险行为往往通过一系列看似合理的步骤累积产生。这一观点挑战了传统的安全评估方法，暗示我们需要关注代理的行为序列而非单一操作。
AI Agent 可以通过标准 MCP 协议直接读取和操作 𝕏 平台：搜索推文、发帖、查看用户信息、管理书签、收发私信等。
大多数人认为社交媒体平台会严格限制第三方自动化操作以防止滥用，但作者指出xAI全面开放了MCP协议支持，允许AI Agent直接执行各种操作，这与主流平台的封闭趋势形成鲜明对比。
Cephalosporins or extended-spectrum penicillins are commonly used (eg, cephalexin, 0.5 g orally four times daily for 7–10 days; see Table 35–6). Trimethoprim-sulfamethoxazole (two double-strength tablets orally twice daily for 7–10 days) should be considered when there is concern that the pathogen is MRSA (see Tables 35–5 and 35–6). Vancomycin, 15 mg/kg intravenously every 12 hours, is used for patients with signs of a systemic inflammatory response.
cephalexin, dicloxacillin, penicillin VK, amoxicillin/clavulanate, or clindamycin (for penicillin-allergic patients). [1-2] These beta-lactam antibiotics provide excellent coverage against streptococci and methicillin-susceptible S. aureus (MSSA
According to agent-centered theories, we each have both permissions and obligations that give us agent-relative reasons for action. An agent-relative reason is an objective reason, just as are agent neutral reasons; neither is to be confused with either the relativistic reasons of a relativist meta-ethics, nor with the subjective reasons that form the nerve of psychological explanations of human action
Agents follow the ReAct (“Reasoning + Acting”) pattern
An Agent often requires multiple calls to the LLM (Thought, Action, Observation, Thought, etc.) to complete a task. Each call incurs cost and latency.1
Tools give agents the ability to take actions. Agents go beyond simple model-only tool binding by facilitating: Multiple tool calls in sequence (triggered by a single prompt) Parallel tool calls when appropriate Dynamic tool selection based on previous results Tool retry logic and error handling State persistence across tool calls
When you bind tools directly to a Model, the model makes a single, stateless decision. It suggests the best tool for the immediate prompt and then stops.
The Agent, however, uses its loop (often ReAct: Reason, Act, Observe) to execute complex strategies
An LLM Agent runs tools in a loop to achieve a goal. An agent runs until a stop condition is met - i.e., when the model emits a final output or an iteration limit is reached.
The difference lies in autonomy and execution flow: A Model with Tools (via direct binding/function calling) is a single, stateless step where the LLM merely suggests the best tool and its arguments, requiring the developer to manually execute the tool and initiate any subsequent calls. In contrast, an Agent with Tools leverages an Agent Executor to manage a dynamic, multi-step loop (e.g., ReAct), where the LLM acts as the planner, deciding which tool to call next, and the Executor automatically runs the tool, feeds the observation back to the model, and repeats the cycle until the complex, multi-step goal is autonomously achieved.
what happens in living beings uh in living organisms
for - adjacency - TAG - living systems - cell agents - consciousness agent
tag the t agent framework t
for - definition - TAG - Tame AGent framework
What is an agent? read more in detail
if you go to another culture and you don't go through the participatory transformation, right? If you don’t, and you're just experiencing culture shock - domicide - the agent arena relationship isn't in place! Then none of those other meaning systems can work for you. There'll be absurd. They won't make sense. That's what he means by it being a Meta-Meaning system.
for - adjacency - culture shock - example of domicide - when the agent-arena relationship is not in place - participatory knowing - meta-meaning system - source - Meaning crisis - episode 33 - The Spirituality of Relevance Realization - Wonder/Awe/Mystery/Sacredness - John Vervaeke
Historically, AI was a tool
for - quote - AI: from tool b to agent - Roman Yampolskiy
quote - AI: from tool b to agent - Roman Yampolskiy - (see below)
Agent
Agent架构
by 2027 rather than a chatbot you're going to have something that looks more like an agent and more like a coworker
for - AI evolution - prediction - 2027 - AI agent will replace AI chatbot
Inpractice, an agent is an app running on a smartphone
Agent = process
the Bodhisattva vow can be seen as a method for control that is in alignment with, and informed by, the understanding that singular and enduring control agents do not actually exist. To see that, it is useful to consider what it might be like to have the freedom to control what thought one had next.
quote: Michael Levin
comment
example - control agent - imperfection: end
triggered insight: not only are thoughts and actions random, but dreams as well
According to the Bodhisattva model of intelligence, such deconstruction of the apparent foundations of cognition elicits a transformation of both the scope and acuity of the cognitive system that performs it.
Adaptive Stress Testing with Reward Augmentation for Autonomous Vehicle Validation
```js // Log the full user-agent data navigator .userAgentData.getHighEntropyValues( ["architecture", "model", "bitness", "platformVersion", "fullVersionList"]) .then(ua => { console.log(ua) });
// output { "architecture":"x86", "bitness":"64", "brands":[ { "brand":" Not A;Brand", "version":"99" }, { "brand":"Chromium", "version":"98" }, { "brand":"Google Chrome", "version":"98" } ], "fullVersionList":[ { "brand":" Not A;Brand", "version":"99.0.0.0" }, { "brand":"Chromium", "version":"98.0.4738.0" }, { "brand":"Google Chrome", "version":"98.0.4738.0" } ], "mobile":false, "model":"", "platformVersion":"12.0.1" } ```
```idl dictionary NavigatorUABrandVersion { DOMString brand; DOMString version; };
dictionary UADataValues { DOMString architecture; DOMString bitness; sequence<NavigatorUABrandVersion> brands; DOMString formFactor; sequence<NavigatorUABrandVersion> fullVersionList; DOMString model; boolean mobile; DOMString platform; DOMString platformVersion; DOMString uaFullVersion; // deprecated in favor of fullVersionList boolean wow64; };
dictionary UALowEntropyJSON { sequence<NavigatorUABrandVersion> brands; boolean mobile; DOMString platform; };
[Exposed=(Window,Worker)] interface NavigatorUAData { readonly attribute FrozenArray<NavigatorUABrandVersion> brands; readonly attribute boolean mobile; readonly attribute DOMString platform; Promise<UADataValues> getHighEntropyValues (sequence<DOMString> hints ); UALowEntropyJSON toJSON (); };
interface mixin NavigatorUA { [SecureContext] readonly attribute NavigatorUAData userAgentData ; };
Navigator includes NavigatorUA; WorkerNavigator includes NavigatorUA; ```
Ownership – identifiers created must be able to have their management restricted to particular agent;
{Single Agent}
a collection of autonomous agents
Agent Based Modeling
// Insight Maker is used to model system dynamics and create agent based models by creating causal loop diagrams and allowing users to run simulations on those
[Neumann, Gros, NeurIPS, 2022] - "SCALING LAWS FOR A MULTI-AGENT REINFORCEMENT LEARNING MODEL"
Page recommended by @wfinck. Seems @karlicoss is the author. This project seems similar to what I've been trying to do with Hypothes.is, Obsidian, Anki, Zotero, and PowerToys Run but goes beyond the scope of my endeavors to just quickly access whatever resource comes to mind (without creating duplicates). The things that Promnesia adds beyond my PKM stack is the following: - prioritize new info - keeping track of which device things were read and how long
Right? You said... No, no, bullshit. Let's write it all down and we can go check it. Let's not argue about what was said. We've got this thing called writing. And once we do that, that means we can make an argument out of a much larger body of evidence than you can ever do in an oral society. It starts killing off stories, because stories don't refer back that much. And so anyway, a key book for people who are wary of McLuhan, to understand this, or one of the key books is by Elizabeth Eisenstein. It's a mighty tome. It's a two volume tome, called the "Printing Press as an Agent of Change." And this is kind of the way to think about it as a kind of catalyst. Because it happened. The printing press did not make the Renaissance happen. The Renaissance was already starting to happen, but it was a huge accelerant for what had already started happening and what Kenneth Clark called Big Thaw.
!- for : difference between oral and written tradition - writing is an external memory, much larger than the small one humans are endowed with. Hence, it allowed for orders of magnitude more reasoning.
A cognitiveagent is needed to perform this very action (that needs to be recurrent)—and another agent is neededto further build on that (again recurrently and irrespective to the particular agents involved).
This appears to be setting up the conditions for an artificial cognitive agent to be able to play a role (ie Artificial Intelligence)
CSP is taking away too much of the user's power and control over their browser use
Whether to inject behavior into a Web page is my choice. How I do so is nobody's business. If a need that can be met with a bookmarklet instead requires a set of browser-specific extensions, that's a tax on developers.
browser extension
I've spent a lot of time in frustrated conversations arguing the case for browser extensions being treated as a first class concern by browser makers (well, one browser maker). But more and more, I've come to settle on the conclusion that any browser extension of the sort that Wildcard is should also come with the option of using it (or possibly a stripped down version) as a bookmarklet, or a separate tool that can process offline data—no special permissions needed.
(This isn't because I was wrong about browser extensions; it's precisely because extension APIs were drastically limited that this becomes a rational approach.)
Krueger, P., Callaway, F., Gul, S., Griffiths, T., & Lieder, F. (2022). Discovering Rational Heuristics for Risky Choice. PsyArXiv. https://doi.org/10.31234/osf.io/mg7dn
To be clear, I am not advocating overthrowing the state or any of the other fear-mongering mischaracterizations of anarchism. I am advocating ceasing to spend all of our resources focusing on the state as the agent of the change we seek. We no longer have the time to waste.
Karimi, Fariba, and Petter Holme. ‘A Temporal Network Version of Watts’s Cascade Model’. ArXiv:2103.13604 [Physics], 25 March 2021. http://arxiv.org/abs/2103.13604.
Gupta, Prateek, Tegan Maharaj, Martin Weiss, Nasim Rahaman, Hannah Alsdurf, Abhinav Sharma, Nanor Minoyan, et al. ‘COVI-AgentSim: An Agent-Based Model for Evaluating Methods of Digital Contact Tracing’. ArXiv:2010.16004 [Cs], 29 October 2020. http://arxiv.org/abs/2010.16004.
This is a useful little tip.
Gordon, D. E., Hiatt, J., Bouhaddou, M., Rezelj, V. V., Ulferts, S., Braberg, H., Jureka, A. S., Obernier, K., Guo, J. Z., Batra, J., Kaake, R. M., Weckstein, A. R., Owens, T. W., Gupta, M., Pourmal, S., Titus, E. W., Cakir, M., Soucheray, M., McGregor, M., … Krogan, N. J. (2020). Comparative host-coronavirus protein interaction networks reveal pan-viral disease mechanisms. Science, 370(6521). https://doi.org/10.1126/science.abe9403
Smaldino, Paul E., and Cailin O’Connor. ‘Interdisciplinarity Can Aid the Spread of Better Methods Between Scientific Communities’. MetaArXiv, 5 November 2020. https://doi.org/10.31222/osf.io/cm5v3.
Identify your user agents When deploying software that makes requests to other sites, you should set a custom User-Agent header to identify the software and provide a means to contact its maintainers. Many of the automated requests we receive have generic user-agent headers such as Java/1.6.0 or Python-urllib/2.1 which provide no information on the actual software responsible for making the requests.
Gleeson, J. P., Onaga, T., Fennell, P., Cotter, J., Burke, R., & O’Sullivan, D. J. P. (2020). Branching process descriptions of information cascades on Twitter. ArXiv:2007.08916 [Physics]. http://arxiv.org/abs/2007.08916
Sturniolo, S., Waites, W., Colbourn, T., Manheim, D., & Panovska-Griffiths, J. (2020). Testing, tracing and isolation in compartmental models. MedRxiv, 2020.05.14.20101808. https://doi.org/10.1101/2020.05.14.20101808
Marshall, B. D. L., & Galea, S. (2015). Formalizing the Role of Agent-Based Modeling in Causal Inference and Epidemiology. American Journal of Epidemiology, 181(2), 92–99. https://doi.org/10.1093/aje/kwu274
allows you to deploy "'strict-dynamic' in a backwards compatible way, without requiring user-agent sniffing
Each request to the API must be accompanied by a user agent request header. Typically this should be the name of the app consuming the service.
sentinel
An agent tracking developments in human-specified, or agent-sensed, topics. Like a watchlist, google search, etc, It would need to be able to do topic mapping and merging.
Check a landlord or agent
A computer simulation of refugees’ journeys as they flee major conflicts can correctly predict more than 75 percent of their destinations, and may become a vital tool for governments and NGOs to help better allocate humanitarian resources.
(like “nature itself,” notmerely our representations of it!) has a history
RE: "Nature itself" having a history
Nathaniel's in-class comments last week were very helpful to hear prior to the readings this week. It is particularly helpful to consider the paradox that some folks want to protect the Earth from humans and somehow return it to a point "before humans," as though the Earth exists outside of humans and we are pure agents acting upon it. (Which is where we get things like this video that has been going around Facebook because of Earth Day:) https://www.youtube.com/watch?v=49w7GHVYoI0
This video pretends the Earth is an agent, but it actually only reflecting human actions back on humans. There is an underlying argument that our relationship to the planet is only the things we do to it and not all the other relationships and existences on and in it.
That changed with the end of WWII. Waves of discharged soldiers subsidized by the GI Bill, joined by the children of the expanding middle class, wanted or needed a college degree.