YAML is the workflow layer for AI agents

The next unlock in AI agents probably isn't a bigger model. It's a YAML file.

That sounds like a letdown until you notice where the time actually goes. The bottleneck isn't reasoning. It's setup tax. Antoine Rousseaux made this case cleanly: before an agent does anything useful, you pay a tax to prepare it.

Every new session, you re-paste context. You re-explain your process. You remind the agent about tone, constraints, the examples that matter, and the last five things it keeps forgetting. Then, finally, you ask it to do the work.

Skill bundles kill that tax. Group the skills and standing instructions under a single-named workflow, and the ritual disappears.

We've been living this with Majestic Marketplace. Every capability pack is a pack.yml, a portable manifest that declares what the workflow exports:

name: marketing
description: Marketing research, messaging, content, and discoverability guidance.
version: 0.1.3
depends_on: []
exports:
  skills:
    - brand-positioning
    - content-writer
    - hook-writer
    - linkedin-content
  prompts: []
  references:
    - content-operations
    - brand-and-positioning-systems
  scripts: []
targets:
  codex: supported
  opencode: supported
  pi: supported

That's not just config. That's a workflow product. The workflow is the product, and YAML is the packaging.

Why it matters:

• No setup tax. The agent loads the right skills, references, and instructions without another context-paste ritual.
• Portable by default. The same pack model can target Codex, OpenCode, Pi, and whatever runtime shows up next.
• Version-controlled process. Your workflow lives in git, not in a clipboard or a buried chat history.
• Composable. A pack can be small and focused, or lean on another pack when the capability is actually needed.
• Runtime adapters stay secondary. The portable workflow is the source of truth. Runtime-specific glue is optional.

Here's the part that many AI SaaS companies will underestimate. The SaaS can still be the data store. But the owned YAML file becomes the workflow layer, and that's a very different product boundary than the one most tools are defending.

One caveat: this only pays off if your workflows actually recur. A genuine one-off doesn't need a manifest, it needs a prompt. The moment you catch yourself re-explaining the same process twice, that's the signal to package it.

Most people will keep re-onboarding their agent every morning. The ones who compound will build a library of named workflows and run them like tools.