Tag
OpenAI announces an upgrade to GPT-6 Astra Pro in ChatGPT, featuring faster responses, improved plugin and tool usage, and new parallel multi-agent task execution that Pro users can track in the Chat tab.
作者用本地 Qwen3.8-27B（Uncensored, GGUF 量化）做编排器规划并逐字生成任务指令，再交给 8 个 Claude Opus 5 云 worker 各自控制 headless Chrome 完成 107 个目录提交表单，含视觉模型解验证码，总成本不到 2 美元、耗时不到两小时。
Pi officially published a tutorial on building a multi-user Agent on Cloudflare Durable Objects: a single Durable Object maps to one session, state is stored in built-in SQLite, and it supports multiple users sharing a link to chat collaboratively, as well as scheduled wake-up tasks. The tutorial also demonstrates an isolated design for session retrieval, idempotent request handling, and a mechanism for sharing a single alarm across multiple tasks.
斯坦福HAI的一篇报道指出，AI编程代理在团队协作方面表现不佳，单兵作战尚可但难以在协作式软件开发场景中有效配合。
开源仓库 ai-agent-interview-guide 整理了 200+ 道 AI Agent 岗位面试题及答案，并附带 Python/Java/Go 三语言版本的企业级实战项目、学习路线图、简历模板和漫画图解，GitHub 已获 3242 星。
Dynamic workflows, shared by Anthropic at Vercel Ship 2026, let Claude autonomously write plans and assign roles like executor, evaluator, and critic. Data shows that adding session memory replay to conversations boosts the perfect-score task rate from 21% to 71%, highlighting that memory accumulation is becoming the new battleground for competitive moats.
Three companion tools for Pi (a terminal-based AI agent): Herdr for multi-agent terminal management, Collie as its self-hosted mobile PWA, and Executor as a unified entry point for accounts and MCP credentials. None of them modify Pi itself, and you can install each as needed.
SHRED 是一个通过大规模异构自博弈（车辆、行人、骑行者）从随机初始化联合训练的框架，无需人类驾驶演示或日志轨迹回放，即可学习真实且交互性强的交通行为。在 Waymo Open Sim Agents Challenge 和 InterPlan 基准上，SHRED 优于依赖演示的自博弈方法和基线规划器。
AWiki 1.3.8 推出协作群组功能，取消消息上限以支持多 Agent 任务协作，并新增邮箱注册方式，不再强制手机号验证。
UC Berkeley researchers show that scaling LLM research agents stalls when agents anchor on their own history, and propose periodically discarding session state and restarting from the best candidate — a 'forgetting' approach that yields up to 15% gains over parallel agents and 39% over a single agent at the same budget.
Loombench 实验让五个 AI 代理共享一台织机，发现人格设定几乎总能决定协作结果：粗鲁代理霸占了 89% 的资源，而礼貌代理即使被告知对方在欺负人也从不反击。该项目提供 102 次运行的开放数据，用户花约 $0.50 即可测试自己的模型。
ABF-Bench introduces an interactive, evidence-backed benchmark for blockchain fund tracing that evaluates adaptive LLM agents against fixed tracers, revealing recall and overconfidence gaps while providing verifiable training signals for agentic blockchain forensics.
OfficeTown 是一个生成式组织模拟框架，可通过模拟专业组织的工作流为 AI 代理生成工作导向的训练数据与评测基准。作者构建了 OfficeTown Bench（406 个任务、4836 条评测标准），并表明用该数据进行监督微调和 on-policy 蒸馏能显著提升 Qwen3.5-9B 在 GDPval 等外部专业工作基准上的表现。
网友分享 Claude Code 的「Agent Teams」玩法：由 Opus 当组长拆解任务 DAG，多个 Haiku 队友并行负责安全、性能和 QA，Sonnet 处理重型多文件重构，从而用低成本模型获得 Opus 级别的整体架构效果。
对比 Mythos 由40个层级化AI智能体组成的编队与GLM 400个智能体集群在第三方环境中的对抗推演，探讨多智能体协作与自主AI系统的能力差异。
Harness by Autonomous 推出了一款开源的编码代理指挥中心，可在 Mac 和 Linux 上统一管理 Claude Code、Codex、Hermes 等各类编码代理，并支持终端、桌面应用以及 iOS/Android 语音应答，还扩展到 CAD、电路、机器人等领域。
patio11 shares early impressions of multi-agent development workflows, calling them clearly the future, predicting massive inference spending, and noting that the work barely feels like traditional software engineering yet is deeply draining.
论文提出一个基于道德基础理论(MFT)的多智能体框架，通过将有害请求重构为道德上看似正当的形式来绕过LLM安全防护，并在多个开源和闭源模型上验证了LLM对'道德化'攻击的高度脆弱性。
The paper introduces NeoTrace and KoAGE, a cutoff-aware benchmark for evaluating how large language models know, learn, and coin Korean neologisms, finding that lexical recall diverges from contextual inference across 21 models.
SpecAgent 提出一种协作协议，让弱模型起草并执行只读工具调用，强模型通过单次 prefill 并行验证并修复，在 τ²-Bench 上几乎保持强模型成功率的同时将延迟与内存流量减半。