The Model Is Rented. The Harness Is What You Own.

What a harness actually is

The model is the engine. The harness is the rest of the car.

Every agent that does real work inside a business is wrapped in a layer the model vendor did not build. That layer decides five things.

What data the agent can see, and where that data comes from. Not “connected to Salesforce” in the abstract, but which objects, which fields, and whether anyone trusts them.

What systems it can touch, and what it is allowed to write back. Reading an opportunity is one thing. Changing the close date is another.

What it can do without a human signing off. The line between “draft the email” and “send the email” is a policy decision, and someone has to make it.

What it remembers from one task to the next, and from one session to the next, so it is not starting from zero every morning.

How you know it got the answer right. The tests you run before it acts, and the log you check after.

That is the harness. The model is rented from Anthropic, OpenAI, or Google, and will be swapped out twice before your next renewal. The harness is specific to your business, your data, and your rules. It is the part you own. Building it well is what people are starting to call harness engineering.

We were already doing this. We called it AI hardening.

I will be transparent. When I heard the word in New York, my first reaction was not “what a new idea.” It was “that is what we have been doing inside Salesforce for a while now, and we gave it a worse name.”

We called it AI hardening. The name told clients what we were protecting against: an agent that updates the wrong record, a prompt that leaks a field it should not see, a workflow that runs fine in a sandbox and falls over on real data. What the name did not say was what we were building. We were building the harness. The permissions model, the data contracts, the approval steps, and the tests that let a client say yes to an agent touching production.

The reason that work takes months instead of days is boring. It was never the model. It was the data underneath it and the rules around it. Which brings me to the part of this I think matters most.

Single player to multi player

Here is the frame I have landed on for where most companies actually are.

Single player AI is one person, one chat window, one prompt at a time. It is genuinely useful. It makes a good analyst faster and a mediocre writer passable. But it leaves nothing behind. Nobody else can see what happened, nobody can audit it, and when that person leaves, the prompts leave with them.

Multi player AI is agents running inside the business on shared data. They act on records other people depend on. They have permissions, an audit trail, and a way to tell when they are wrong. That is what “production AI” means in practice, and almost nobody is there yet.

The gap between single player and multi player is not a smarter model. The models are already good enough. The gap is the harness. Every company I talked to in New York that had actually shipped an agent into production had, whether they used the word or not, built one.

The three roads

Companies now have a real choice about where that harness comes from, and the choice is more consequential than it looks.

The first road is the enterprise harness. Salesforce announced theirs in September. ServiceNow and HubSpot are building the same thing. If your work already lives inside one of those platforms, this is the path of least resistance. The permissions model is already there. The data is already there. The governance story writes itself for the CIO. The catch is in the next section.

The second road is a third party harness that sits across your systems. A growing set of vendors sell a layer that connects to Salesforce, HubSpot, your warehouse, and your email, and lets you build agents on top of all of it. More flexible, faster to stand up, and not owned by any one of your platform vendors. The trade is that you now have one more system of record to govern, and the integration quality to each source system varies more than the demo suggests.

The third road is building your own. I assumed this was reserved for companies with hundreds of engineers. It is not. Several large enterprises have built their own harnesses to run inside the organization, and I do not think they will be the last. The cost of the agent loop itself has collapsed. What remains expensive is the part that was always expensive: the data and the rules.

I am not going to tell you which road is right. It depends on where your data lives and how much of it you trust. Which is the real question.

The boring part decides the outcome

Whichever road you pick, the harness is only as good as the data arriving from the source systems.

This is the uncomfortable truth underneath the whole conversation. A capable agent on a messy CRM does not fail loudly. It does not throw an error. It gives you confident wrong answers at scale. Duplicate accounts become two different customers. A stage field nobody updates becomes a forecast. A contact role that was wrong for three years becomes the person the agent emails.

Every production AI project I have seen succeed started with an honest look at the data, and most of the budget went there. Every one I have seen stall skipped that step because the demo looked fine.

Does this mean you need a warehouse?

Here is the question I kept asking in New York, and the answer surprised me.

Does a data warehouse, or at least a deliberate data strategy, accelerate production AI over the long term? Several companies told me they had already decided. They are solidifying their data strategy outside their source systems. Not because Salesforce or HubSpot cannot hold the data, but because they do not want any single platform to own the context their agents depend on.

That is the part of the enterprise harness decision I would think hardest about. Pick the platform harness and you get speed and governance on day one. You may also be picking who owns your business context for the next decade. The vendor’s harness will always see the vendor’s data best. If your agents need to reason across Salesforce, your product database, your support tool, and your finance system, somebody has to assemble that picture, and the companies ahead of the curve have decided that somebody should be them.

I do not think a warehouse is a prerequisite for a first agent. Start small, inside the system you already have, and earn trust. But I do think the companies that treat their data layer as a strategic asset, rather than something that lives wherever the application vendor put it, will move faster on the second, fifth, and twentieth agent. The harness is where that compounding happens.

The takeaway.

The model is rented. The harness is what you own. Production AI is a harness engineering problem, and the hardest part of the harness is not the agent. It is the data feeding it and the rules governing it. Decide early who owns your context, because the harness you choose is making that decision for you.