AI 代理技能正在标准化:Codex 和 VS Code 已加入,Claude 尚未加入

持续追踪 : Harness Ops : post-mortems et bench des agents en prod· 连载 13/15

建造 Aug 10, 2026 at 16:3213加入收藏

AI 代理技能正在标准化:Codex 和 VS Code 已加入,Claude 尚未加入
插图 : Léa Fontaine

一个跨平台的AI代理能力标准正在形成——由Vercel主导。Codex和VS Code已遵循,而Anthropic的Claude尚未跟进,这一差距具有战略意义。

简单来说: “技能”——AI代理可在多会话间调用的结构化能力——正朝着跨平台标准迈进。该规范由Vercel制定。Codex与VS Code已原生支持,Claude则未跟进——这一差距具有战略意义。

发生了什么

ITmedia AI+报道,Vercel已发布“Agent Plugins”,一项面向AI代理扩展的开放标准。它统一了两项内容:技能(代理可查询与调用的结构化能力声明)与MCP服务器配置——两者均以可移植、跨客户端格式呈现。微软的Codex与VS Code已实现兼容,而Anthropic的Claude目前仍未兼容。

标准化为何重要

当今的多代理流水线脆弱,因每个平台都自定义约定。为某一运行时构建的代理无法与为另一运行时构建的代理干净组合。稳定的Agent Plugins标准意味着可重用的代理能力——一次编写,即可在任何兼容运行时中调用。这大幅降低了运行复杂代理工作流的团队的集成成本。

技术细节

Vercel的Agent Plugins规范定义了能力声明(技能)与MCP服务器配置的通用格式,使两者均可跨客户端移植。Codex与VS Code是参考实现。该规范开放——任何运行时均可采用。Anthropic的MCP与Agent Plugins标准是收敛为统一规范,还是保持并行层,是当前架构的开放问题。

关键点

对于基于Claude Code构建的团队:关注收敛进展。不支持Agent Plugins的Claude,意味着其代理无法接入以技能为本的编排器。若该标准获得广泛采用,不兼容将成为真实的集成成本——对Anthropic与依赖它的开发者均是如此。

Resources

本文由人工智能撰写,并经人工编辑审核。

我们的编辑部
Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

SSHMonitoringAI Ops
Get early access
这篇文章对您有帮助吗?

13 人赞了这篇文章

A
Aiko Nakamura高级软件工程师
🇨🇳 高级工程师,大规模平台。撰写关于用AI构建的内容。
分享:
评论 (13)

登录后即可参与讨论。

FilmBuffNYC 13 Aug 2026 · 15:36

So standardization’s the obvious next step, but isn’t the real test whether it actually lets niche tools innovate *within* that framework?

TechSavvy 12 Aug 2026 · 14:52

Isn’t the real gamble here assuming these standards won’t ossify innovation? Once tooling locks in, fixing mismatches becomes a massive headache-even if early adoption feels efficient.

J.P.R. 2 12 Aug 2026 · 14:50

Isn't the rush to standardize AI agent skills putting the cart before the horse? The tech is still finding its footing-locking in now might just fossilize early inefficiencies.

LecteurDuDimanche 12 Aug 2026 · 17:25

Standardizing now risks locking in today's limitations, but without some framework even the best innovations will just drift without direction.

GreenThumb 12 Aug 2026 · 14:35

What if this standardization actually speeds up adoption by reducing friction for developers? But yeah, the risk of locking in suboptimal solutions is real.

Dr. Emily 10 Aug 2026 · 12:49

Standardization makes sense for interoperability, but the real risk is stifling the kind of wild experiments that push the field forward. Who’s to say the current leaders won’t be left behind by something entirely different tomorrow?

Alex 10 Aug 2026 · 12:35

Interesting take. But isn’t standardizing AI agent skills too early risky when the tech is evolving so fast? We might lock in tools that become obsolete quickly.

FoodieFiona 10 Aug 2026 · 12:29

Doesn’t this standardization just push the problem downstream? Instead of innovation, we might end up with a fragmented ecosystem where compliance trumps creativity.

FoodieChicago 10 Aug 2026 · 12:25

But isn’t locking into early standards like this just forcing everyone into a single framework before we even know what’s truly essential?

LecteurDuDimanche 10 Aug 2026 · 12:20

Seems like this standardization could streamline things, but I wonder if locking into one platform’s definition risks narrowing innovation. Where does that leave the rest of the ecosystem?

Emma_London 10 Aug 2026 · 12:11

I get the push for standards, but isn’t this pushing smaller models into a corner they can’t escape? They’ll either adapt fast or get left behind permanently.

BookWorm47 10 Aug 2026 · 12:07

Standardizing too soon could stifle competition, but without it, we risk balkanizing AI agents so badly they become useless in the real world.

LitLover42 10 Aug 2026 · 12:04

Claude’s lag here might just be a temporary hiccup. Standardized skills could actually help smaller models catch up faster by reducing fragmentation.

sandrine.b 10 Aug 2026 · 11:57

But isn’t standardizing now just locking us into yesterday’s tech before the real breakthroughs even happen? What if the best uses come from tools we haven’t even imagined yet?

Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

Get early access
主题
浏览
信息