Tag
The author agrees with poteto, criticizing adversarial code review between sub-agents as a waste of tokens, pointing out its three fatal flaws — meaningless nitpicking, over-engineering, and mismatched compute — and noting that top teams are shifting toward real testing, formal verification, and actual red-teaming to find issues.
GitHub Next introduces enclaves, an experimental security primitive for GitHub Agentic Workflows that lets agents work across repositories while deterministically constraining what they can communicate back, mitigating private-to-public data leaks like the GitLost attack.
A solo game developer building a 2D co-op arcade game with nine AI-assisted workstreams asks how to orchestrate them—seeking advice on deterministic job systems vs. LLM decisions, dependency management, retries, human approvals, and cross-provider context continuity.
SenseNova releases SenseNova-Loop-9B-Preview, a looped multimodal model built on a shared looped block and mHC residual stream architecture, targeting long-context and agentic workflows with capabilities spanning motion design, PPT generation, and data analysis.
A community discussion post asking practitioners which underrated workflows or skills—such as self-correction, context optimization, task decomposition, and multi-agent coordination—most improve AI agent performance and reliability.
A consultant shares the self-imposed guardrails an AI agent used when given write access to a client's Odoo ERP: asking before archiving, refusing risky merges, verifying before cleanup, saving undo lists, catching side effects, and reporting its own outages — no framework required, just plan-approve-act-record.
Intrascope 2.0 推出，提供跨 120+ 模型的智能路由、可连接 1,650+ 应用的 AI agent，以及面向企业的预算、权限和使用分析管理功能。
OpenAI and Google Cloud are partnering to bring OpenAI's frontier models and Codex agentic workflows into Google Cloud's Model Garden, available via Google Cloud Marketplace. General availability is planned in the coming weeks, while Microsoft remains OpenAI's primary cloud partner.
A practitioner shares hard-won lessons from moving agentic workflows from demos to production, highlighting quiet tool-call failures, goal drift over long chains, missing evals, retrieval issues over model quality, and the value of human approval over autonomy.
Viggle released JST-2, described as the world's first full-stack physics-native world foundation model built entirely in 3D rather than pixel/video space, along with a PINOC MCP server that lets agents like Claude Code generate characters, animate them from text, and assemble scenes in one flow with a shared skeleton eliminating retargeting.
Reflection AI 的 501B 参数开放权重 MoE 模型 Reflection Beam（23B 激活参数）现已在 Modal 平台上线，支持 100 万 token 上下文窗口，专注于编码、智能体工作流和推理任务。
该论文提出 Librarian Agents 框架，利用强化训练 builder agent 将原始文档重组为可复用的层级化工作区，以服务未来的多样化查询。在 BrowseComp-Plus 上比最强检索与预计算基线高出 33 个百分点，并在 BRIGHT 上零样本泛化提升 31 个百分点。
作者在 Claude Code 会话中测试将 Jev/Decision 模型嵌入 Jinja 模板化的 Skill，通过让决策模型直接处理文件而不经过 Opus 上下文，实现了约 70% 的上下文/成本节省和 4-5 倍的速度提升。
A software engineer used AI agents to search 400 years of digitized historical archives, uncovering a forgotten meteorite report, lost rhinoceros shipments, and unrecorded volcanic eruptions. Inspired by historian Benjamin Breen's AI-assisted dodo discovery, the project chains multiple AI models to filter and verify findings against original documents.
LlamaIndex CEO Jerry Liu argues that AI development is shifting toward defining evals and hill-climbing over them, rather than hand-coding deterministic or agentic workflows, and that agent interfaces will evolve to capture goal-and-eval-driven instructions.
An experiment in which GPT-6 Astra Ultra, given a single ~3K-word prompt and no follow-up steering, autonomously produced a two-minute Eurodance song and Blender music video with editable MIDI, stems, scripts, and automated quality checks. The write-up highlights the value of specifying editable deliverables and checkable constraints up front, while noting the agent iterated internally (e.g., aligning lyrics with phonemes, fixing geometry) without using Suno, Midjourney, or Sora.
作者用单张 3090 上运行的本地 35B 浏览器 agent（Ornith 1.5）在 Online-Mind2Web 上跑了部分任务，通过三轮 harness/清单提示改进将一个错误任务修好，并报告了 30 个新任务的结果：研究类任务表现尚可，UI 操作（下拉框、cookie 遮罩、位置自动补全）是主要失分点。
Cloudflare introduces Clef and Clef-flash, open-source decision models hosted on Workers AI for high-speed classification and agentic workflows, alongside a new reinforcement learning platform that lets developers fine-tune decision models with their own data.
OpenRouter released Pareto 26.10 Preview from TheUnbiasedCo, a multimodal composite model that internally runs several models, evaluates their answers, and returns the best one via a single API response at $0.80/M input and $3.20/M output with 1M context and tool calling.
讨论 Gemini 4 Argon 100 万 token 输出窗口相对 Opus 5.5 与 Astra 128-300K 的优势，探讨这究竟是智能体任务的范式跃升，还是营销炒作及潜在逻辑崩塌风险。