Somewhere in Lower Manhattan, an AI is doing its tenth-thousandth pivot table.
It is 11:47 p.m. on a Tuesday in 2026. A few floors above a coffee shop in New York, a screen flickers with what looks like Salesforce. It isn't. It's a replica of Salesforce, every dropdown faithful, every quirk preserved, every loading spinner punctually annoying. Inside the replica, an AI agent is trying - again - to update an opportunity stage without breaking the parent record. It fails. It tries a different click. It fails. It tries something a human would never try, and somewhere in a neighboring office, a researcher watches, leans back, and says, almost to herself, "There we go."
This is Fleet AI. Not a model. Not an app. A gym. Specifically, a gym where the next generation of software agents lifts the digital equivalent of weights, eats the digital equivalent of broccoli, and learns the difference between confidence and competence.
The picks-and-shovels of the agent economy.
If 2024 was the year of the foundation model and 2025 was the year of the demo, 2026 is the year someone finally asks the obvious question: where, exactly, do agents learn to do the work?
Fleet's answer is unglamorous and very lucrative. Agents learn the way humans do - by repetition, by feedback, by being wrong inside a room where being wrong is cheap. Fleet builds the room. The room looks like Excel. Or Salesforce. Or whatever software your team actually uses on Wednesdays. Inside that room, models can fail at a million tickets before they ever touch a real one.
The customer list reads like the guest list of an AI dinner party: frontier labs, hyperscalers, large enterprises that have noticed their existing data is not enough. They don't want a chatbot. They want an agent that arrives on day one already fluent in the company's stack.
Numbers that matter.
Nicolai Ouporov has a thesis about work.
Nic Ouporov, co-founder and CEO, is the kind of person who can describe the labor market and a reinforcement-learning loop in the same sentence and have both make sense. His thesis is simple, and he says it often: humanity is moving from doing work to directing it. Soon every person will be a manager. The catch is that the new hires will be agents, and the agents will need training.
The team he assembled looks less like a startup and more like a stage of "Ocean's Eleven" for AI: alumni of Anthropic, xAI, Meta Superintelligence, Essential AI, Contextual AI, Mercor, Docker, Citadel, Jane Street, Cruise. The Cruise alumni are the tell. Cruise spent years building simulators for self-driving cars before the cars ever drove themselves. Fleet is doing the same thing for software.
Three products. One bet.
Training Gyms
Faithful replicas of the software teams use - Salesforce, Excel, internal tools. Agents practice. Humans supervise. The reward signal is real because the environment is real.
RL Environments
Bespoke reinforcement-learning environments built for frontier labs, hyperscalers, and enterprises. The hard data that doesn't exist yet, manufactured to spec.
Drop-In Assistants
Agents who arrive at a customer already fluent in the customer's stack. Less "set me up" and more "ready, what's first."
How we got here.
Engineers, researchers, and artists.
Fleet describes the team as engineering-first, but the public bios mention researchers and artists, which is a small detail and a large signal. The product is, at root, a believable world. Building believable worlds is what artists do. Building them in code is what engineers do. Doing it at the granularity of real software is what researchers do.
The vibe, from the outside: quiet lab, loud customer base. Few interviews. No splashy launch video. The kind of company that lets the round speak first.
On camera.
It is 11:48 p.m. The pivot table works.
One floor up from the coffee shop, the AI updates the opportunity stage correctly. The parent record holds. The researcher writes a single line of feedback and the model takes the lesson home. Tomorrow, a different agent will fail in a different replica of a different piece of software, and someone will watch, and write another line.
The thing about gyms is that no one talks about them at the medal ceremony. You hear about the gold and you hear about the athlete, and you almost never hear about the room with the wooden floors. Fleet is fine with that. Their bet is that in a decade the most important sentence about modern work will be: the agents that do it were trained somewhere. And the somewhere will be a few floors above a coffee shop in New York, where a screen, mistaken for Salesforce, taught a generation of software workers how to work.