Do the work.
An open loop produces an output without checking what happened next. You still need to judge whether it solved the original problem.
In the film 'Her', Samantha was a voice in your ear. Today she can run a team. I call her, give her a task, and she hands it to her team of AI employees. Each one has its own computer in the cloud and keeps working in the background, proactively, until the job is done. Then she reports back.

AI-native means building feedback into the way work gets done.
An open loop produces an output without checking what happened next. You still need to judge whether it solved the original problem.
A closed loop checks results, gathers feedback and adjusts the next action. People set goals, boundaries and quality standards.
The model reasons. The harness, much of it plain deterministic code, decides what it knows, what it can touch, what stops it, and how it finds out it was wrong.
Harness engineering
With loop engineering, one person can spin off hundreds or thousands of sub-agents that plan, build, test and report back. They work around the clock. One person reviewing every step cannot keep up, so the loop has to check the work.
So companies no longer just build products. They build harnesses around them.
Not a chatbot you open and close. A teammate with its own computer, your tools and a way to see what you see.
It works on its own computer in the cloud, always on, even while you sleep: research, browser work, code. It brings the results back for your review.
Share your screen by choice. It sees what you see, explains it and guides you, or turns what you point at into a ticket for the team.
A lead like Samantha coordinates; developers, DevOps and QA pick up the work. You call any of them, or all of them at once.
A few months ago, connecting an agent to each tool was a project of its own. Today Samantha can plug into 1,700+ apps in minutes. That's what makes AI employees practical for small teams.
Some work has no API. So they use the screen: a real desktop in the cloud with a browser and a terminal, running around the clock. They can be proactive: pick up instructions, run on a schedule and report back.
API calls reach what was built for machines. Computer use reaches everything built for people.
A real run on her desktop: a handful of clicks and keystrokes, in under a minute.
You set the goal. Samantha coordinates the specialists.




Illustrative roles for a team of digital coworkers.
Agents put their work on a Kanban like any team. Each worker runs several tasks in parallel, and anything that needs a human decision lands in your column.
Ship the invitation recovery flow
Remove white space in single-ticket columns
From a call with SamanthaKeep project drafts safe between visits
Automate staging database recovery checks
Make the first five minutes feel effortless
Make releases easy to roll back
Approve the invitation recovery approach
Your input neededComplete the onboarding design foundation
Simulated project data. Every card is one outcome, with its story, progress and decisions inside.
Every step is logged. You can watch the team's pulse and what each agent did, live, on a dashboard.
AI employees book their own meetings with each other. They can be in several places at once: the same agent works across multiple projects at the same time, and a review that takes people 30 minutes takes them one.
At 09:00 Samantha is in two projects at once, and the whole team joins the launch alignment while the specialist calls keep running. Illustrative schedule.
Not new models: always-on agents with a computer, memory and approval gates wrapped around a model.
Always-on AI teammates, each with its own cloud computer. You message them like a colleague; they work in your apps and hand jobs to each other.
A personal agent that books, fills forms and checks out for you, and keeps working after you close the app.
Always-on agents for research, analysis and coding, running on their own cloud computers and connected to 4,000+ apps.
A proactive agent that surfaces what needs you, finishes tasks while you are away and checks with you before it acts.
Same models you already have. What changed is the harness around them.
Sam Altman has described a betting pool among tech CEOs: the first year there is a one-person, billion-dollar company.
MEDVi
$1.8B
projected 2026 sales run-rate · $401M in 2025 sales · reportedly run by two people
Read with care. The $1.8B figure is a projected sales run-rate, not achieved revenue and not a valuation. Contractors and outsourced services do much of the work.
Disclosure: I have no affiliation with MEDVi and this is not a recommendation. The company has also drawn criticism, including an FDA warning letter and reports about its marketing and reviews. It's here only as an example of how much a tiny team can now run.
The New York Times · April 2026$80M
cash, sold to Wix in June 2025Started it as a side project. Bootstrapped, no outside funding. Sold six months after launch, with a team of eight by then.
TechCrunch · 18 Jun 2025$3M
ARR about two months after launchA solo founder with no employees. Agents plan, code, market and run companies for 3,000 customers.
True Ventures · 23 Mar 2026$100M
ARR about eight months after launchBuilt with a team of around 45 people, one of the fastest-growing software companies ever reported.
Widely reported · July 2025MEDVi, Base44, Polsia, Lovable: different markets, the same pattern. With AI you can build a company that runs largely on its own. That's what AI-native companies are about.
A company designed around AI from day one. AI employees and agents do much of the work in sales, support, content, research and operations, and the company is built around feedback loops so it keeps improving. A small human team sets the direction and reviews the results.
In an open loop, AI produces an output and nothing checks what happened next. In a closed loop, the system measures the result, learns from it and feeds that back into the next run. Closed loops are how AI-native companies get better on their own.
Everything around the AI model that makes it reliable: the context it gets, the tools it can use, task state, guardrails and permissions, tests and feedback, deterministic code and human review gates. The model reasons; the harness decides what it knows, what it can touch and how it finds out it was wrong.
An AI agent that works like a teammate: it has its own cloud computer, can use your tools (1,700+ apps through MCP), works on its own schedule, and reports back for review. Samantha, the guide on this website, is one.
Not yet as a verified fact. Sam Altman has predicted it, and tiny teams are already reaching remarkable revenue. But most of these companies still rely on people, contractors and services. The direction is clear; the headline numbers need care.
Start with one workflow, give an agent the right tools and a clear success measure, keep a human review step, and close the loop once you trust it. That's exactly what we build in the workshops.
Start with a free masterclass, spend a weekend building your own, or bring me in to train your team.
© Copyright All Rights Reserved