Skills Aren't Replacing Agents. They're Naming the Harness.

A YouTube Short going around makes a sharp claim: stop building agent frameworks, just use Anthropic’s Skills. One general agent (Claude Code), a folder per workflow with SKILL.md and script.py, done. The author calls it the secret most-powerful tool: a bunch of folders and markdown files.

He’s right about the architecture. He’s wrong about the vocabulary, and the vocabulary is doing real damage.

The video keeps the word “agent” intact and just argues you only need one of them. That misses the deeper move. The reason this architecture works isn’t that one agent beats seven. It’s that there is no agent at all — there is a harness, and the word “agent” was wrong from the start.

What’s actually happening

Forget the homunculus. The picture the word “agent” paints — a little worker sitting inside your computer, waiting, deciding, possibly scheming — is why people are simultaneously terrified of “agentic AI” and underwhelmed by what their agents do. The mental model is wrong, so the reactions are wrong.

The better frame:

You don’t run agents. You harness intelligence.

A harness is what you put on a horse or a dog. You’re not creating a worker. You’re taking something already powerful and pointing it at a narrow job for a finite time. That’s the whole thing.

Once that lands, the rest of the architecture falls out:

1. You run a routine, not an agent.
2. The first run orchestrates — it async-launches more harnessed thinking.
3. Each launched instance is harnessed by a skill: data, instructions, tool access. That’s the bridle.
4. Inside the harness, the intelligence gets a narrow job.
5. Job done, the harness terminates and releases the intelligence. The API call closes.

It is not sitting around dormant. It is not pacing the cage trying to find something to break. It existed for the duration of a task and then it was gone.

So what is a skill?

A skill is the harness. The folder of markdown and scripts the video celebrates isn’t a side dish next to the agent — it’s the bridle that turns raw intelligence into a worker that does one thing safely. SKILL.md is the instructions. script.py is the tool access. The data is the field of view. Strap that on a base model and you have a harnessed instance, ready to do a job and disappear.

That’s why “one Claude Code + many skills” outperforms “seven agents.” Not because one agent scales better than seven, but because skills correctly name the thing doing the work. The orchestrator isn’t seven workers; it’s a routine that bridles intelligence on demand. Every skill invocation is a harness opening, doing a job, closing.

Why the word still matters

Two consequences worth naming.

On safety. Most “agent risk” arguments quietly assume the homunculus model — a thing with continuity, motive, drift. A harnessed instance has none of that. It has a job and a kill-switch built into its shape. That doesn’t make every system safe, but it does mean the conversation should be about how the harness is built, not whether the horse is plotting.

On building. If you are designing a “general agent that can do anything,” you have already lost. Generality lives at the orchestrator. Every leaf does one narrow thing under a tight bridle and then disappears. The work is in the harness — the data you give it, the instructions you write, the tools you scope. That’s the craft.

The video’s instinct is right: the secret tool is folders and markdown. But what those folders are matters. They aren’t a workaround for not having a real agent framework. They’re the architecture that makes “agent framework” the wrong abstraction in the first place.

The intelligence is a commodity. The harness is the product. Skills are the harness.

Stop calling them agents. Build the bridle.