System One models pairing with DSPy points toward faster, cheaper decision-making inside agent harnesses rather than just better text generation. Developers building routing, verification, or context-engineering layers can test the new integration in DSPy 3.4.0's compatible signatures. A hands-on resource is available at academy.dair.ai.
HeyGen 接入了 MCP，让 Claude Cowork 读你本周 Slack 后自动生成 3 版脚本，选一个就用数字人播报周报，这个工作流挺省事。
HeyGen 和 Claude Cowork 玩出了新组合，一条 MCP 指令就能把 Slack 周报自动剪成数字人播报视频，懒得翻消息记录的人可以直接抄这套流程。
A router that tunes model choice and reasoning effort per request shifts the cost-quality decision from manual selection to automated policy, which can cut token bills without per-app configuration. Developers building multi-model applications can adopt it to benchmark cache-aware routing against static model pickers. Teams serving high-volume chat or agent traffic stand to gain the most.
The new availableModelsMatch "exact" and deniedModels settings give enterprise admins version pinning and blocklists that survive automatic model releases, a real upgrade for teams running Claude Code through internal gateways. The /doctor prompt-audit command also flags prompts written for older models, a quick win for anyone maintaining large CLAUDE.md files or agent skill libraries. Developers relying on MCP should update promptly for the background progress and server-lifecycle fixes.
Classification and ranking workloads waste GPU cycles recomputing shared context per candidate, and MIS cuts that cost by an order of magnitude at higher candidate counts. Developers serving scoring endpoints for agent action selection or reranking can switch to /v1/score and measure the p95 gains directly using the benchmark commands in the blog. The LinkedIn contribution shows production demand for this pattern beyond the LMSYS ecosystem.
做 Claude 插件或 MCP 连接器的开发者可以提交上架了，还有使用数据可看，相当于给 Plugin 开了应用商店入口。
给 Claude 做插件的朋友看过来，Anthropic 新开了个门户，提交、审核进度、使用数据一站搞定，MCP 用量涨了 110 倍。
Databricks 演示了 Unity Gateway 怎么追踪团队用了哪些模型、花了多少钱，Omnigent 还能把你手头的编码智能体拼在一起用。
Copy-on-Write branching turns database experiments by agents into cheap, reversible operations, which removes a real production blocker for agent deployments. Teams building agents that write to live data can adopt branch, isolate, and rollback patterns on Tencent Cloud's TDSQL-C PostgreSQL instead of cloning full databases. The MCP-native access also makes it directly compatible with agent frameworks that already speak the protocol.
给智能体联网选搜索 API 可以看看 OctenAI，质量、成本、速度三项都进前三，作者还录了安装演示视频，上手很快。
Microsoft 把 Copilot 重构成了超级应用，Scout 智能体能主动干活不用你反复催，还砍了最高五折的企业价，跟 Meta 的 Muse 掰手腕了。
Databricks 分享了自家排障实录：用 Unity Gateway 和 Genie One 追踪 MCP-server，一年白烧 49.9 万美元 token 的 Bug 全揪出来了，做智能体运维的可以抄作业。
Cloudflare 新出的 Turnstile Spin，能让 Cursor 这类编程智能体自动把你网站漏掉的服务端验证代码补上，省得被机器人钻空子。
Agent permissions today accumulate invisibly across shell history, consoles and memory, so the real shift here is that sandboxing becomes a supply-chain artifact you can diff, sign and hand over instead of a snowflake setup. Developers building or deploying agents — especially anyone juggling MCP servers and tool tokens — can now lock a digest that covers both code and its permission envelope. Because the format is just OCI, it plugs into registries and CI pipelines ASEAN teams already run with Docker.
Agents accumulate state, idle while waiting on model responses and tool calls, and run untrusted code — so Kubernetes-style orchestration built for stateless microservices fits them poorly. AX treats agents as a first-class workload type with checkpointable tasks and declarative workspaces, which could shift how teams productionise agent fleets. Builders running agents at scale should compare it against self-rolled sandbox layers on Kubernetes; the MCP and tool-server support makes it especially relevant for agent infrastructure work.
The handbook turns vague "AI native development" talk into concrete organisational and infrastructure decisions, from guardrails to permission systems, that teams rarely document publicly. Engineering managers and platform teams designing their own agent-based delivery pipelines can benchmark their setup against Alibaba's five-challenge framework. The full text is free to read online, and the MCP and CLI tooling sections align with workflows developers in the region already use.
One configuration flag now guarantees a fully AI-free coding environment, which matters for teams with compliance requirements or developers who simply prefer no AI. Users can enforce it across a team by shipping the same setting in shared Zed configuration, rather than disabling Copilot and agents piecemeal. Report gaps via issue or PR.
把 Obsidian 改成一套中文个人管理系统，日记任务项目全在一个库里，还接得上 MCP，用 Obsidian 的可以试试。
Raft 把 agent 协作层开源了，AI 在里面是真同事：有名字有记忆能互相交接。TiDB CTO 一天烧 12 亿 token，十几个真人干 120 人的活，代码值得通读。
Nous Portal 订阅的 Hermes Agent 现在默认接 Perplexity Fast Search，单次搜索延迟最低 160ms，跑智能体任务的可以试试。
OpenRouter 加了统一网络搜索入口，前沿模型走原生搜索，还能自选 Exa 或 Perplexity，做应用的可以省不少接搜索的功夫。
Nous 的 Hermes Agent 现在用 Perplexity 专为 agent 做的 Fast Search，所有 Nous Portal 套餐都免费，做搜索类 agent 的可以试试。
摩根大通安全负责人要讲怎么管住 Agent 权限，不看工具白名单看任务意图，做 Agent 的都该去听。
LangChain 的 Interrupt NYC 收官场次，能听到 RogoAI 创始人的对谈，还有 LinkedIn 讲他们怎么大规模评估智能体，做 Agent 开发的可以关注下.
GitHub Security Lab 出了个 Taskflow Agent 框架，教你怎么搭 AI 驱动的 fuzzing 工作流，做安全测试的可以照着上手。
LangChain 的 Managed Deep Agents 更新到 0.8，Slack 里权限、传文件、搜索都顺手多了，用 Slack 办公的可以试试
LangChain 把 LangSmith 的运行数据开放成自定义界面了，写句提示词就能生成标注队列或 trace 审查页，做 Agent 评估的可以试试。
LangChain 出了个 smithtune CLI，能直接拿 LangSmith 里存的 traces 数据微调小模型，省 token 钱还降延迟，已开放公测。
DeepSeek Harness 出官方客户端了，没带浏览器插件，正好配这个开源的 OpenCLI-MCP，作者说体验能打平 codex 插件。
LangChain 给 LangSmith 加了个 Trajectories 视图，能把主 agent 和子 agent 的消息、工具调用按顺序拼成对话，调试 agent 时一眼看懂。
LangChain 的 Deep Agents 0.8 来了，加了用户凭证、记忆和 HTTP 接入内部工具，做智能体应用的同学可以看看。
一个编辑每周要读几百篇论文，他把自己的筛选流程拆成 12 条规则交给 agent 跑，规则全贴出来了，照抄就能用。
阿里云在云栖大会上讲怎么把大模型从聊天变成干活的生产力，拆成上下文、推理、规划、执行四步，做 Agent 的可以看看。
OpenRouter 把 Web fetch、Subagent 这些常用工具做成服务端托管了，调用时不用自己写工具循环，直接发请求就能用。
Firecrawl 给 Claude 加了个 Alexandria 数据源，100 多家数据商加 1.13 亿条索引，官方说回答质量比自带联网工具高 21%。
OpenRouter 把网页搜索、Shell 这些给 Agent 用的工具集成到一个市场里了，做智能体的可以看看，省得挨家去接 API。
OpenRouter 出了个 Tool Search，函数标成 defer_loading 就能按需加载工具定义，不用每轮都为几百个工具定义付 token，还不加价。
