Harness

Every major AI lab is quietly telling you the model isn't the product.

Look at what they've shipped in the last twelve months. Anthropic built Claude Code and Claude Cowork — harnesses for developers and knowledge workers. OpenAI shipped Codex, Projects, and Frontier. Google shipped Gemini CLI. They've all seen the same thing from the inside: a chat window isn't enough, and the next tier of value lives in what wraps the model.

That's the piece most business owners are still missing.

The model is not the product

When you open ChatGPT, you're not using "AI" in some abstract sense. You're using OpenAI's chosen presentation of the underlying model — their personality, their guardrails, their pre-decided context limits. It's capable, but it's a chat window. It has no hands. It can't pull last month's figures from your accounting system. It can't update a client record. It can't watch your inbox.

Claude is the same story. Different personality, different defaults — same underlying constraint.

For the model to do anything operational, something has to wrap around it. That something is the harness.

The model is the brain. The harness is the body.

The six pieces of a harness

A harness has six parts. Each one is a deliberate design choice.

1. Perception. It can read your inbox, scan your documents, watch your data. Input from the systems where the work actually lives.

2. Action. It can update a record, send a message, write a file. Not describe the action — do it.

3. Event response. It triggers on state changes. A new enquiry arrives. An invoice goes overdue. A threshold is crossed. No prompting required.

4. Persistent memory. It remembers your clients, your context, what happened last time. Not starting fresh every conversation.

5. Domain training. It's calibrated to your specific operations. Your terminology, your decision criteria, your quality thresholds.

6. Authorisation. It knows what it may do alone, what needs approval, when to escalate. Boundaries built into the architecture, not bolted on.

Most deployments use three or four. The point is that each is a deliberate design choice — not a bundled feature on someone else's platform.

What this means for you

If the question in your head is which AI should we use?, you're asking the wrong question. It's like asking whose brand of pipe cutter you should buy when the real question is whether you have plans for the house.

The right questions are harness questions:

• What does the system need to perceive?
• What should it be able to do on its own?
• What events should trigger it?
• What does it need to remember across time?
• What does it need to know about your specific operation?
• What may it do alone, and what needs sign-off?

Answer those six and the model choice answers itself. Probably whichever one you can access cheapest today.

A good time to be a tinkerer

Two years ago, building a custom harness meant hiring an agency for six months and paying half a million dollars. Today, a couple of people with API keys can build something genuinely useful in six weeks.

The models got smarter, but that's not the story. The tinkerer's tools caught up — frameworks, evals, skill libraries, agent runtimes — and the cost of harness-building collapsed with them.

If you're running a business with 10–40 people and you've been watching the AI headlines feeling vaguely behind, here's the thing: you don't need to pick the right model. You need to build the right harness.

Your business is the harness. Everything else is commodity.

Karl Howard · Reforged · 22 March 2026