Brett Queener, a partner at Bonfire Ventures, uses “harness” at the scale of an entire product in his newsletter Tales from The Bonfire:

The Horse is the LLM model: enormously powerful, fundamentally unpredictable, not something you command so much as ride.

The Harness is the application layer fitted to that animal for one rider and the entire job that rider does, holding the ontology, the interaction model, and the learning loop.

The Hay is everything the horse and harness need to stay healthy at scale: orchestration, governance, cost routing, security, observability, data infrastructure.

The horse is the AI model. The hay is the collection of services that keeps it running. The harness is the product someone uses for the whole job. It holds what the system has learned about that job: the facts, relationships, exceptions, and rules Queener calls its “ontology.” It also decides which steps need the user’s attention and uses previous results to improve the next attempt.

The developer harness keeps its instructions, tools, and files stable while models change. Queener applies the same principle to an AI business. A company can use cheaper models for routine tasks and save the most capable models for difficult judgment. This keeps its cost of goods sold (COGS) low and protects its profit margins:

First, no smart harness maker will wed themselves to one model provider. Every good CTO I talk to has already abstracted the model away from the product. Not out of loyalty. Out of survival: which model is best changes every 60 to 90 days, and prices move by an order of magnitude.

Portability is also a margin discipline, and it’s most of the difference between the companies quietly running healthy gross margins and the ones complaining in public about COGS. A large share of what runs in production never touches a frontier model: classification, extraction, routing, summarization, retrieval. Reserve the frontier for the hard, ambiguous, high-stakes judgment and send everything else to a pony. Run every call to the biggest model because it’s easier and your gross margin belongs to somebody else’s price list.

The graph describes how jobs and results move through the system. Queener shows that the same graph can serve two roles. An off-the-shelf tool for coordinating any AI system is hay. Once a company fits that coordination to its own agents, rules, and way of working, it becomes part of the harness.

But coordinating the work is not enough. The harness must retain what it learns even when larger customers reach it through another interface, such as Claude, instead of a screen provided by the harness itself. It must also learn from results. Queener calls the tests that score its work “evals.” The system that feeds those scores and corrections back in so future behavior changes is the reinforcement learning, or RL, platform:

It owns the interaction model for that role. Not a chat box, and not necessarily a UI at all. What the rider is asked, what they approve, what they never have to look at again, what escalates and how. Which doesn’t require pixels: upmarket this increasingly gets consumed headless, rendered inside Claude or something the customer built on Vercel. Serve that, but treat headless demand as a scoreboard rather than an architecture, because yours is the surface holding the ontology and the loop and the memory of this rider, so its ceiling should be higher than a general assistant’s at your rider’s own job. The test is what survives a change of surface. Swap the front end next year: if everything they taught the system is still there, you were the harness. If not, you were endpoints with good docs.

It owns the eval and RL platform. The dividing line is whether a graded outcome changes future behavior without an engineer deciding what to change. Everybody has dashboards, and a human reading one and then writing code is a feedback process, not a loop. The test: can you point at a behavior your product has today that nobody on your team designed?

And it is singular. One rider, one harness. Nobody runs three of these at once any more than they drive to work in two cars. Not “fewer apps,” but one, per rider, per job.