The lab
Cultivator AI is an AI lab based in Eindhoven. Our software factory engineers build agentic harnesses for companies. We measure what they deliver and publish what we find.
Three things: a company's knowledge, structured so that every agent gets the right context; the harness that runs its processes with agents; and the factory that builds this for the next company faster than for the last.
Everything that now sits in heads, emails and spreadsheets, in one place. You send a note, a quote or a photo of the whiteboard to the brain and it lands where it belongs: who signs above a certain amount, which customer always gets priority, why that one quote is built differently. You ask for it back from your phone and get the answer with its source. What it delivers: new people are up to speed sooner, and the question that always sat with one person no longer costs half a day.
This is where it all comes together: the harness is the system that runs your processes with agents. An order arrives as an email, the agent reads it and enters it in your ERP with process-specific knowledge attached, and your people decide where it matters, from their own working environment. It runs on your own servers and you own it: if you stop working with us, it keeps running. Ours is called Trellis, and Cultivator AI itself runs on it today.
The factory is what makes it fast. Every company needs its own harness, around its own processes. Built by hand this takes too many AI engineers, so we build a harness factory. A harness is the main interaction layer for steering agents. Central, so there is no sprawl of loose tools and agents. And sovereign where it is needed.
Three steps, for every build: elicitation, build, publish. At each step you see what it costs you and what it brings, in hours and in numbers you can check yourself.
01We sit with the people who do the work, in the room where it happens. They show how the process runs today: who decides, which rules apply, where the exceptions are. That becomes a list of what the agent needs and a baseline: how many hours the process costs you now. You know up front what we will build and what we will measure it against.
02We build the agents for that one process, coupled to your systems, sovereign where it matters. Your people watch from their working environment and decide at every step that matters; the agent puts the doubtful cases in front of them. Within thirty days it runs on your real work.
03We measure before and after: hours, lead time, error rate, and the share of cases the agent handles without a person. You get those numbers with date and source, including the ones that did not go our way, and you decide on them whether the harness stays and which process is next. What we learn goes on this page, without your name and without your figures, unless you want them there.
Measurements on our own company, which is customer zero. Each one says when, and where the number comes from. There is nothing from a customer here, because there is no customer yet.
Latest measurement
443 / 504
From 15 September to 15 October, 443 of the 504 changes merged into our own platform were built by our builder agents and reviewed by an open-weight model running on our own hardware.
Source: the titles of the merged changes on our platform's main branch, read on 15 October 2026.
An open speech model runs on our own machine faster than real time: in 8-bit, every 80 milliseconds of audio costs 70 milliseconds of compute, the answer comes after about 0.6 seconds, and a laptop on our own network adds 4 milliseconds. Earlier that day, before running it in 8-bit, it was too slow for a conversation.
Source: our own measurement on the machine, 12 October 2026.
We measured three ways of reading whether our AI reviewer is raising a real objection, on 472 held-out reviewer reasons. Our existing rules missed 79 percent of the real objections, a small hosted model missed 31 percent, and a local open model missed 92 percent. The whole measurement, 777 questions, cost 1.7 dollar cents. We continued with the hosted model.
Source: our own measurement report of 11 October 2026.
Our five growth agents made their first real model calls: seven, all seven passed every automatic check, for about 0.06 dollar cents together. Read by a person, two were unusable: one wrote a customer story that never happened, one wrote the message as if the prospect were the sender. Both went back to the builder. A passed check is not a read text.
Source: the receipts of the run on our own server, 9 October 2026, and the by-hand read of 11 October.
Our LinkedIn program runs on the same harness: agents draft, a person approves and posts by hand, about two hours of our time a week. The posts point here, not the other way round.