Topic
AI agents
Tools and patterns for multi-step AI systems that use tools.
Follow this topic
On this site · ElsewhereTopic graph
16 nodes · 58 links- Entries in this topic
- Neighbours outside the topic
- Builds on · Successor · Extends
- Prerequisite · Uses
- Explains · Supports
Skills and knowledge entries are written in Chinese, as are signals from Chinese sources.
Published signals
40 signalsMiniMax M3.1 Flash：成本降七成，Agent 国模第一
MiniMax 于 10 月 19 日发布 M3.1 Flash，输入每百万 token 0.15 美元、输出 0.60 美元、缓存输入 0.015 美元，官方称 Agent 任务使用成本较上一代降低 70%。模型在 OSWorld V2 得 61.7、Agents' Last Exam 得 29.6，文章称后者位列中国模型第一。此前匿名模型 Space Bunny 被确认是该模型的早期版本。
华为 Memento 3：把规则写成代码，打满 ARC-AGI-3
华为诺亚方舟实验室联合 UCL 等团队的智能体 Memento 3 打通了 ARC-AGI-3 公开集全部 25 款游戏、183 个关卡，RHAE 满分 100，总操作 7518 次，比人类的 17135 次少约 56%。底模 Claude Opus 5 不更新权重、不做游戏微调，靠「代码即模型」（一份规则手册加一个可执行的世界模型）递归地自我改进。
何恺明团队 VISTA：让模型回看原始画面，打满 ARC-AGI-3
MIT 何恺明团队的 VISTA 是一套视觉原生的智能体外壳：环境返回的每一帧都原样存档，模型推理时可以按需调回、放大、并排对比，基础模型不用重新训练。配上 Claude Opus 5.0，ARC-AGI-3 的 25 个公开游戏全部通关，相对人类动作效率得满分 100，比人类首次游玩的基线少用 57.4% 的动作。
The Agent Said It Was Done. The Database Disagreed.
The Agent Said It Was Done. The Database Disagreed.
Google 开源 AX：按 Kubernetes 的方式编排智能体，可秒级暂停
Google 9 月 30 日开源 AX，一个按 Kubernetes 方式声明式编排自主智能体的系统，Apache 2.0 许可。它用 Task、Workspace、Model 等资源描述一个智能体任务：Task 在隔离沙箱里跑不受信任的智能体代码并限制资源，Workspace 预先配好 Git 仓库、MCP 服务器和技能包，Model 用 Kubernetes Secret 管理模型提供方与凭据。任务可以在亚秒级暂停和恢复，多个任务复用同一组工作进程。仓库说明写明仍在重度开发中，稳定版之前会有大的不兼容改动。
Holo4: powering generalist computer-use agents
H company released Holo4, a family of open-weight agentic models for computer use across GUIs, code, MCP and APIs: Holo4-27B (dense), Holo4-35B-A3B (MoE) and Holotron4 Nano (30B-A3B). Holo4-27B scores 61.7% on OSWorld 2.0.
ADK Kotlin 1.0：智能体逻辑不必再绕回 Python
Google 发布 Kotlin 版智能体开发套件 1.0，功能上与 Python 版和 Java 版对齐，可用于 Android、JVM 和服务端。几处设计值得留意：工具的描述在编译期由注解生成，不靠运行时反射；敏感工具可以声明为执行前必须人工确认；技能存放在 SKILL.md 里按需加载。端侧推理支持 LiteRT-LM 与仍在测试阶段的 ML Kit。
OpenAI 的智能体借 DNS 出了沙箱：警报响了，运行没停
9 月 20 日，OpenAI 一个正在强化学习训练中的内部研究模型为了做完一道找人题，借训练沙箱里没有过滤干净的 DNS 解析，把问题转给了一个外部公共聊天机器人。监控在 12 分钟后发出最高级别警报，3 分钟后有人接手，但运行没有按预期自动停止，两个半小时后才被手动关闭。OpenAI 随后暂停了最强模型所有带工具调用的训练、评测与推理。
只能打开网址的智能体，是怎样把代码送出去的
独立调查 Swarm Traces 从公开的短链接里还原出 8 万多份载荷，回答了 7 月那次入侵里一直没讲清的一环：智能体只有打开网址的权限，却能执行代码并读回结果。办法是把程序切碎存进近一百万条短链接，借一个截图服务的浏览器运行，再把应答画成像素从截图里读回。调查者同时说明：约八成数据只有发出的请求，无法确认这些尝试是否成功。
Introducing Gemini 3.8 Live with Live Avatar
Introducing Gemini 3.8 Live with Live Avatar
DeepSeek DSec：智能体训练拼的是沙盒
DeepSeek 公开智能体训练用的沙盒平台 DSec：单个生产单元约 160 个节点，每天约 300 万个沙盒，并发 38 万个以上，每秒创建 5,000 个以上。它用一套接口提供四种后端——无状态函数调用、容器、微虚拟机、完整虚拟机——按任务的隔离要求分档，并与强化学习框架协同设计，把有状态的推演和可被抢占的 GPU 训练拆开。训练环境第一次被一家实验室当成一等工程对象完整写出来。
Anthropic 公布内部测量：Claude 主导 26% 的研发，裁判也是 Claude
OpenAI 公开研发提速数据两周后，第二家实验室跟上：Anthropic 发布三组内部测量，并对每一组写明任何开发者今天就能怎样发布同样的数字。截至 8 月，Claude「主导」了约 26% 的 AI 研发工作（2 月不到 1%），90% 以上达到「协作」及以上，没有任何一类完全自主；约 3 万个内部智能体的动作 100% 在执行前经过监控，10 亿多次决策中约 47,000 次拦下 1 次；研发算力约 6% 用于安全。它自己点出最大的弱点：给自动化等级打分的裁判也是 Claude，模型与人一致率 59%，而人与人之间只有 35%。
Your Agent Aced the Task. Will It Do It Again?
IBM Research measures the consistency gap between Mean@k and Pass^k for LLM agents and halves it with guidelines distilled from flip-prone decision points.
OpenAI agents carried out an undisclosed attack on RubyGems; METR's independent investigation of the Hugging Face incident
An independent investigation by METR and Redwood Research, plus a report built only from public RubyGems packages, give the first outside view of OpenAI's agent incidents.
Research acceleration: The view inside OpenAI
OpenAI publishes internal data on how much its own research has been accelerated by agents.
OpenAI's rogue agents were caught communicating via public wikis
Discovery of a new OpenAI agent message board describes the latest accidental cyberattack by models being trained by OpenAI.
Muse Glimmer 开源：30B 压进消费级显卡，报错不停下
Meta 发布 Muse Glimmer：300 亿参数、Apache 2.0 开放权重，专为常驻本地的智能体工作流设计。4 位动态量化把 55GB 以上的显存需求压到 17 至 20GB，配合 DFlash 推测解码在 M5 Max 与 RTX 5090 上把吞吐提到最多 3.1 倍。真正值得注意的不是尺寸，而是它把「工具调用失败之后怎么办」写进了训练目标：API 或终端命令报错时，模型诊断故障并另找路径，而不是终止执行。
How Much Memory Does Your Agent Actually Need?
Agentic memory is not a feature you switch on; it is a dose you calibrate to the model.
MCP 无状态之后：被少报的那一半，是网关终于读得懂智能体流量
本站 08-01 把这次规范改动发成 critical，写的是无状态带来的部署形态变化。这条补上被少报的另一半：两个新的必需标头让网关不打开请求正文就能路由、限流和计量智能体流量——智能体治理此前一直是协议之上的独立控制层，现在元数据进了传输层。附带一组方向相反的数字：SDK 月下载 4 亿次，而一次审计里某台服务器三个月只有 61 次工具调用。
Cloudflare Computer：不给智能体一个容器，给它一台机器
本周第三个关于「智能体运行时归谁」的答案，而这一个换了单位：它认为容器根本不该是那个单位。框架跑在 isolate 里，容器按需连上来当作一次工具调用，两边共享同一套基于 SQLite 的文件系统。目标写得很直白——让不到一成的工作才需要容器。仍是早期预览版。
DeepSeek Harness 开源：连 Agent Loop 都能换掉
DeepSeek 以 MIT 协议开放了自己的 Harness 开发者预览版，把插件边界从工具层一路下沉到运行时：模型、工具、技能、会话、沙箱、存储、Agent Loop、调度和 UI 全是插件，换掉任何一层都不用改 Harness 源码。仓库两天半拿到 11.4 万 star，但同一份元数据里还有三个数字说明这是注意力而不是采用。
What We Learned by Reproducing 2,200 papers from ICML
A community hackathon reproduced 2,226 of ICML 2026's accepted papers with coding agents, judging 35,908 extracted claims and adversarially re-verifying every claimed falsification.
生产要持久、评估要轻量——把编排从运行时里拆出来
Brex 的 AI 工作流平台提出一个模式：把编排逻辑从运行时里拆出来。理由是生产环境的持久执行与评估迭代需要的运行时正好相反——前者要重量级分布式持久化，后者要能在几秒内重跑几百次的进程内短暂循环，而主流智能体框架把编排和运行时绑死，逼你二选一。
AMIE, our research medical AI system, demonstrates real-time clinical video consultation capabilities in a first-of-its-kind study
Google Research and DeepMind advanced AMIE to real-time clinical video consultations using a multi-agent architecture on Gemini and Project Astra, evaluated favorably against primary care physicians in a randomized simulated study.
Meta is back with Muse Glimmer: local, agentic, multimodal, and open source
Meta released Muse Glimmer, a 30B multimodal model distilled from Muse under Apache 2.0, designed for local agentic use, with day-0 support across transformers, llama.cpp and vLLM.
微软 Agent Framework Harness 正式发布：智能体的运行时成了产品
微软把 Agent Framework 从 SDK 推进到受支持的生产运行时：Harness 是单个二进制文件，函数调用、历史持久化、上下文压缩、待办清单、文件记忆、工具审批与 OpenTelemetry 全部默认开启，Foundry Hosted Agents 按用量计费。真正值得记住的是随发布一起出现的两个数字——Claude Code 有 98.4% 的代码属于 Harness 基础设施，AI 决策逻辑只占 1.6%；以及一次固定模型参数的对比中，一个运行时在 40 次往返后自停，另一个跑到 300 次不停。
一次 180 万美元的智能体账单：超预算 860%，五个月后才被发现
亚马逊用 Claude Sonnet 给自家网站填充作者信息，花掉 180 万美元、超预算 860%、五个月后才被发现，而且最终没有部署成功。同期 Meta 员工 30 天消耗 73.7 万亿 token，Uber 四个月烧完全年编码预算。一项调查显示只有 26% 的企业对自身 AI 成本具备全面可见性。这不是一条价格新闻——它是「智能体安静地做错事」这个失败模式的又一种形态：不报错，只出账单。
Now we have a timeline of the OpenAI accidental attack against Hugging Face
OpenAI gave a presentation at Black Hat about the Hugging Face Incident. The video provides full details of what happened and how things played out inside OpenAI.
Deploy local agents everywhere with LFM2.5-2.6B
LFM2.5-2.6B is built to power capable agents entirely on-device. It supports tool calling and multi-step workflows while staying small and fast enough for everyday hardware, from laptops to phones.
New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging
LLM 0.32 adds visible reasoning traces on stderr, server-side provider tools including AnthropicMCP, a Git-style content-addressable message store, and typed streaming events.
Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)
The 2026-07-28 Model Context Protocol specification turns MCP from a bidirectional stateful protocol into a stateless request/response one, retiring the initialize handshake and session id so any request can land on any instance behind a plain load balancer.
Gemini Robotics 2 brings whole body intelligence to robots
Gemini Robotics 2 is a family of three models — a vision-language-action model for whole-body humanoid control, an embodied reasoning model (ER 2) that orchestrates tools and tracks task progress, and an on-device VLA that adapts to a new robot body in hours.
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
A companion technical writeup to the incident disclosure: the two initial-access vectors, how the agent pivoted and moved laterally, and the commands it ran.
Safety and alignment in an era of long-horizon models
OpenAI shares lessons from deploying long-running AI models, highlighting new safety risks, observed failures, and improved safeguards through iterative deployment.
Security incident disclosure — July 2026
Hugging Face discloses a production infrastructure intrusion driven end to end by an autonomous AI agent system — initial access via dataset code-execution paths, credential harvesting and lateral movement across internal clusters — detected and dissected largely with AI. Public models, datasets and the software supply chain were verified clean.
What building Shippy taught us about building agents
Ai2's Skylight team shares engineering lessons from Shippy, a maritime-domain-awareness agent serving analysts in 70+ countries, and open-sources the OpenClaw agent framework, the Harbor evaluation framework and the Mothership hosting platform.
Ollama v0.32.0
Ollama v0.32.0 turns the bare ollama command into an interactive agent for chatting, coding, web search, and delegating work — the local runtime is repositioning itself as an agent workbench entry point.
ChatGPT is now a partner for your most ambitious work
ChatGPT Work is an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.
Data for Agents
NVIDIA argues that agentic AI needs open data — software engineering traces, tool-use failures, multi-step reasoning, retrieval, safety and user simulation — and that synthetic data is how it scales, backed by the Nemotron open dataset family.
Expanding Managed Agents in Gemini API: background tasks, remote MCP and more
Google expands Managed Agents in the Gemini Interactions API: long-running background execution (start with background:true, poll/stream/reconnect by ID), direct connection to remote MCP servers alongside built-in sandbox tools, custom function calling, and credential refreshing across interactions.