RR

Your AI Agent Doesn't Have a Model Problem

Published: 2026-03-15 · 7 min read

Everybody is asking the wrong question.

They ask, "Which model should I use?"

GPT or Claude. Cheap or expensive. Fast or smart. Latest release or last month's winner.

That question matters some. It is not the main question.

The main question is this: what sits around the model?

Because the model is not the product.

The harness is the product.

That's the part a lot of business owners miss when they start poking around AI. They see a chatbot demo. They see a flashy benchmark. They see somebody on X claiming one model is now 7% better than another. And they assume the magic lives inside the model itself.

It doesn't. Or at least not where the business value lives.

The business value lives in the orchestration, the memory, the rules, the handoffs, the integrations, the follow-up, the error handling, and the way the whole thing plugs into the actual work your business already does.

Av1dlive has used the phrase harness engineering to describe this. That's a good frame. You can see the same principle in systems built by teams like Anthropic and in orchestration-heavy agent stacks like OpenClaw: the surrounding system is what turns a model into an operator.

The harness is the system that makes the model useful in real life.

The model can write. The harness makes it work.

The easiest way to understand it

A truck engine matters.

But if I hand you an engine sitting on the pavement, you do not have a delivery business. You have a heavy, expensive piece of metal.

To make money, that engine needs a truck around it.

It needs a transmission, wheels, brakes, steering, a driver, a route, maintenance, fuel, and a dispatch system telling it where to go.

That is what most AI conversations leave out.

The model is the engine.

The harness is the truck.

And if you're running a business, you do not get paid for owning engines. You get paid for moving freight.

Why smart demos die in the real world

A raw model can do impressive things in a blank text box.

Ask it to summarize a contract. It can do that. Ask it to draft an email. It can do that. Ask it to give you ten marketing ideas. Sure.

Then real life shows up.

Now the contract has to be pulled from the right folder. The email has to use the right customer history. The marketing ideas have to match your offer, your pricing, your market, and your actual voice. Somebody has to check whether the output is wrong. Somebody has to log what happened. Somebody has to decide what to do when the model is uncertain, or when the customer asks a question outside the script, or when it hallucinates a detail that could cost you money.

That is the difference between a cool demo and a useful system.

The raw model is rarely the bottleneck.

The bottleneck is that most businesses don't have a harness.

What the harness actually includes

If the word "harness" sounds abstract, break it into parts.

A real business harness usually includes five things.

1. Orchestration

This is just sequencing.

What happens first? What happens next? When does AI answer directly, and when does it hand off to a human? What information does it gather before it tries to act?

A good AI system doesn't just "know stuff." It follows a process.

If a lead comes in, maybe the harness tells the system to:

  1. identify whether it's a fit
  2. pull pricing and service details
  3. draft the right response
  4. log the conversation in the CRM
  5. schedule follow-up if no reply in 48 hours
  6. escalate if the lead asks about something high-risk

That's not model magic. That's workflow design.

2. Integrations

A model without integrations is trapped in a chat window.

A useful system touches the software you already use.

Your inbox. Your CRM. Your calendar. Your proposal tool. Your document storage. Your phone system. Your billing system.

This is where a lot of business owners get fooled. They think they bought "AI," but what they actually bought was another tab to babysit.

If it doesn't connect to the work, it doesn't reduce the work.

3. Guardrails

This is the part everybody wants to skip until they get burned.

Guardrails are the rules.

What is the system allowed to say? What is it not allowed to say? What requires approval? What gets routed to a human? What counts as a safe answer versus a risky one?

If you run a law firm, a financial practice, a medical office, a home service business quoting jobs, or anything with real customer consequences, guardrails are not optional.

Without them, you're basically letting a very confident intern talk to clients unsupervised.

4. Memory

This is where things start to feel less like a toy and more like an operator.

Memory means the system remembers what happened before.

Not in the spooky sci-fi sense. In the practical business sense.

What did this customer ask last week? What proposal did we already send? Which objections keep showing up? Which leads went cold? Which service tier fits this type of customer? What does this owner hate seeing in a draft?

Without memory, AI starts from zero every time.

That means repeated explanations, repeated mistakes, and repeated friction.

With memory, the system gets sharper over time.

5. Feedback loops

This is how the harness improves.

Where did the system fail? Where did a human step in? Which response got ignored? Which one closed? What kinds of requests keep breaking the workflow?

A decent harness learns from misses.

Not because the model is suddenly self-aware, but because somebody built a process that captures errors and tightens the system.

The local business example nobody talks about

Let's make this concrete.

Say you own a plumbing company in McKinney.

Everybody and their brother is now saying they want "an AI receptionist." Fine. What does that actually mean?

If all you do is bolt a chatbot onto your website, you have not built a business asset. You have installed a polite distraction.

A real harness for that plumbing company would do something like this:

Now we're talking.

The model might help phrase the questions naturally. It might summarize the call. It might handle the back-and-forth.

But the money is not in the wording.

The money is in the harness that routes the right jobs, books them faster, reduces missed calls, and keeps the owner from losing high-value work at 8:47 p.m. because nobody picked up the phone.

That is not theoretical.

That's revenue capture.

And notice something important: if you swap Claude for GPT or GPT for Claude, the value of that harness largely stays intact. You may tune it. You may improve performance around the edges. But the real asset is still the system.

That's why the model is not the moat.

Why business owners should care

Because model quality is becoming a commodity.

The gaps are closing fast. The trend line is clear.

What doesn't commoditize as quickly is a well-built system wrapped around the model.

Anybody can sign up for the same API.

Not everybody knows how to turn that API into a lead capture system, a quoting workflow, a support desk, a sales assistant, a reporting pipeline, or an internal operator that actually saves time and makes money.

That is where the value shifts.

If your AI strategy is basically "pick the smartest model and hope," you're renting intelligence without building infrastructure.

If your AI strategy is "build a harness around the work," you're building an asset.

The mistake small businesses keep making

They start with the tool.

They should start with the bottleneck.

Don't ask, "How do we use AI?"

Ask:

Once you know the bottleneck, then you build the harness around that bottleneck.

That might mean AI is involved in the answer. It might mean AI is only one piece of the answer. Good. That's normal.

AI is usually most valuable inside a system, not standing alone.

What a good harness feels like

A good harness does three things.

First, it makes the model more reliable than it would be by itself.

Second, it makes the business faster without making it sloppier.

Third, it makes the owner's life simpler, not more complicated.

If your "AI solution" requires constant babysitting, ten new dashboards, and a standing meeting just to keep it alive, that is not leverage. That is overhead wearing futuristic clothes.

A good harness is boring in the best way.

It catches the right inputs. It routes work where it should go. It remembers what matters. It flags exceptions. It improves over time.

It quietly makes the business run better.

A note on how I run this

I used to run eight standing agents — each one a named specialist with a permanent role. It worked, but the overhead compounded. Drift, context bloat, tuning debt. Every standing agent is a surface you have to maintain.

I've since moved to two standing agents — an orchestrator and a client-facing operator — with on-demand swarms for everything else. Need a coder? Spin one up. Need a researcher? Spin one up. Task done, agent dissolves. You get the specialization without the standing army.

That's a harness decision, not a model decision. And it's exactly the kind of thing that matters more than which model you picked this quarter.

Where this goes next

The businesses that win with AI will not be the ones with the loudest model opinions.

They will be the ones that build the best systems around the models.

The winners will have tighter intake, better follow-up, cleaner knowledge, smarter routing, stronger memory, and clearer guardrails.

They will not necessarily have the fanciest prompt.

They will have the best harness.

That's the shift.

We're moving from model obsession to system design.

And for small businesses, that's good news.

Because you probably are not going to outspend OpenAI, Anthropic, Google, or whoever else is in the model race this quarter.

But you don't need to.

You need to understand your business better than your competitors do.

You need to know where time gets wasted, where revenue leaks, where handoffs break, where customers wait too long, and where your team is doing work a system should already be handling.

Then you build the harness there.

That's where AI stops being a parlor trick and starts becoming infrastructure.

And infrastructure is where the real value lives.

If you're building agents and want to talk harness design, reach out: deacon@ridleyresearch.com.

The model is not the product. The harness is.

If this was useful and you have questions, email me at deacon@ridleyresearch.com.