The phrase "agents are the new apps" is useful as an analogy, not a compatibility promise. The "LLM as OS, Agents as Apps" paper, by Yingqiang Ge and colleagues at Rutgers, describes a conceptual ecosystem: a model provides a foundation, while agents use tools to carry out particular jobs. It does not establish that every workflow file is a runnable application.
For someone choosing an agent skill, the practical question is simpler: what does this package need in order to do its job?
1. Check where the workflow runs
A reusable Markdown file can describe a job, its inputs and the output to review. It does not by itself supply a runtime, connect an account or install an app. Check the host's supported file formats, tools and account requirements separately.
For example, a campaign-cost review can start with sanitized exports and a written calculation contract. That is different from an agent that reads an advertising account and changes a campaign. Do not infer the second capability from instructions for the first.
2. Check what the standard actually covers
OpenAI says its Apps SDK, available in preview, lets developers build app logic and interfaces inside ChatGPT. It is built on the Model Context Protocol, or MCP, and extends that standard. OpenAI describes portability to hosts that adopt the standard.
That is a conditional claim. It does not mean every agent host supports every app, or that uploading a skill file creates an MCP app. Ask which host, interface, tools and account permissions were tested. Keep an untested host marked untested.
3. Check data access and action limits
A packaged job still needs a clear boundary around the information it can read and the changes it can make. Microsoft's Windows agent guidance describes separate agent identities, scoped access, policy and visibility into agent activity. Those controls are part of the host environment, not something a workflow document can grant itself.
Before using a skill with real records, check what leaves the account, who can see it, how long it is retained and which actions require approval. A useful output contract also names missing facts and stops instead of filling them with guesses.
Treat the analogy as a checklist
The cited sources describe an ecosystem idea, an app-development standard and host controls. They do not establish comparative adoption or prove that one platform is the next Windows. Use the analogy to ask better questions about a specific job, not to skip a compatibility or privacy check.
Start with the complete instructions, run a fictional example, and review the result before supplying authorized real data. A successful example is a readiness check, not evidence that an account connection or real-world action works.
Browse free profession workflows or read the workflow-file guides.