Agent OS and Skills: 26 Years of Tech Trend Insights

Shifting focus from specific technical solutions, the author looks ahead to broader technology trends surrounding agents and skills.

Agent OS and Skills : The author interprets an Agent as an operating system and a Skill as the software that runs on it. Numerous heterogeneous agents can be transformed into Skills that execute on a unified Agent OS, making duplicated agent development potentially wasteful.

Existing mature Agent OSes are grouped into three categories:

Code‑focused Agent OSes: Claude Code and OpenCode dominate this space.

Desktop Agent OSes: Claude Desktop (also called Cowork) and the desktop version of OpenCode.

PaaS/Web Agent OSes: Manus, which currently lacks Skill support, and other platforms that have yet to assume a leading role. A domestic tech giant recently released a 2.0 version of its Agent development platform, attempting to fill this gap but appearing rushed and not fully polished.

Response API trend : Both OpenAI and Anthropic have introduced “response APIs” that expose agentic capabilities via simple HTTP calls and support sandboxed execution. This development suggests that developers may soon rely on calling large‑provider APIs instead of building their own agents from scratch.

Quality‑assured Agent SDKs for building Skill‑compatible Agent OSes include:

Claude Agent SDK – regarded as the most effective but tightly coupled with Claude Code.

OpenAI Agent SDK.

Vercel AI SDK.

GitHub Copilot Agent SDK.

Agent loop and its shells : The author emphasizes that the agent loop is the core of agentic systems. Different “shells” wrap this loop:

Manus adds a web‑based shell.

Coding agents provide a terminal‑UI shell, relying heavily on the Language Server Protocol (LSP).

Desktop agents use a desktop‑client shell.

The distinction among these shells lies in their prompts and tooling. Rather than dismissing the practice of adding shells, the author argues that the diversity of shells enriches the overall agent experience.