All started with chat completions API. Now, companies with the best harnesses are winning. Claude code, Hermes Agent, OpenClaw, Codex, Perplexity Computer, Cursor. And more we don’t know their names yet. They all use similar models, but the harness makes the real difference.
What is a harness?
I was exposed to this word very recently for the first time. And in just a few weeks, it became very meaningful.
First recorded in 1250–1300; it meant "baggage, equipment" in Old French (herneis) and "provisions for an armed force" in Old Norse (hernest). Old Norse is a North Germanic language spoken by Scandinavian peoples roughly from the 8th to the 14th century.
A harness is a set of straps that are put on a horse so it can be hitched to a wagon or a carriage. The harness distributes the effort across larger sections of the horse's body.
Draft horse with traditional harness.
Until recently.
Now it has a totally different meaning, still very relevant. Harness now refers to everything surrounding a model, except the model itself. Can we say the AI models are kind of beasts helping us to gain productivity like in the past, perhaps yes, but we again need harness to get best out of them.
Each AI native company has to have its own harness to survive and become successful in this AI age. I am happy that now things are getting complicated, so at least being able to build an app with one prompt is not a valid argument to founders. If you are building a harness that will be your differentiator, and models are being commoditised but harnesses don’t seem to be. The news came earlier, Cursor launched its SDK opening its harness, it might sound like this thing will be commoditised fast, but it is not like a linear one layered thing, so if you build your harness there are great opportunities to differentiate yourself.
The harness evolved fast
AI companies that once wrapped the models offering light applications are now building harnesses (if they survived) to build deeper and to offer better experiences and create value for their customers.
It started in chat form. Send a question, get a response. Then came function calling, so we could get responses from llms that trigger calls to outer functions like image editing, or so. Then mcp came out so llms could access other tools, then they started to live in our local computers, so could access our files, could setup a small environment to run some python code so they can read our files, create slides and so on. Then they started to create additional files, called memories, keeping track of what is important and some facts so they could remember again. Then they started to access terminal doing all sorts of stuff in our computers. Then small cute files appeared with .md extension, we called them skills, containing neatly sorted instructions so models could generate much better outputs. Then to manage all this stuff, system prompts, tools, memories, skill files, making them available at the right time, they all mattered much more than before.
And the agentic loops started to come to our life, first reasoning came with DeepSeek, all followed, now it is almost default setting in all models, then inside there are multiple llms generating response with different system prompts, and an orchestrator sending responses from one to other, or running in parallel, at the same time tracking the job, to do list etc.
Oh man, this is too much, but I think refreshing for an AI founder since now it’s apparent that it’s not that easy to build a product, we need to build a harness that can differentiate in many ways. We need to make the right judgment where to differentiate, of course having at least a 7 out of ten rating in each part of the harness, but some parts should be 9 out of ten, that can be the winning part. Long-running, persistent, and autonomous, maybe those keywords that will define the next winners.
Harness, once meant military equipment as well.
What we are building
At Quantus Labs, we are building our product Folio, which aims to help millions of businesses run their commerce. Build catalogs, publish stores, transact orders, manage inventory, fulfillment, crm without needing multiple tools. Smart, simple, intuitive in a way that was never possible before. And there are millions of small and medium-sized businesses still undigitized, never heard of Slack or Notion, where many agents trying to connect, running their commerce the old-school way, struggling to adapt to competition due to many inefficiencies adding up in their cost base.
We are building our harness, tools, skills, the right agentic loop necessary for our processes, memory system, alongside all crud operations. It is really fun, and fulfilling to think about the architecture that never existed before, to build a solution that is used, and growing its user base day by day.
If you want to join our team in London or Istanbul, to build the next generation solutions and a unique harness empowering millions, join us.
