Last updated: September 27, 2026
On February 11, 2026, I started an experiment: how far can current frontier models go, and where do they break? Not by bolting AI onto an existing company. By asking a harder question: what does a company look like if humans don't do knowledge or execution work at all? No legacy org chart. No existing team to augment. A clean-sheet design, built from first principles around what the technology actually is.
I started simple, with a content generation loop. That was solvable quickly. So I turned to the harder challenge: engineering. Building commercial software, with real quality standards, for real users in real production environments. Seven months later, I am still pushing. I have not found hard ceilings in current-generation frontier models. I have found model generations that break under this architecture, documented why, and used the findings to make the system stronger.
That's the finding. Not that AI can write code. Not that AI can automate tasks. The finding is that the ceiling everyone is looking for isn't in the models. It's in us. In our ability to design the systems, governance, and organizational structures that turn raw model capability into reliable, repeatable, shippable work.
It took me more than 30 years to figure out how to do this in 7 months. Everything I have learned building commercial software over my entire career went into designing these loops. The patterns for how teams should work, how quality gets enforced, how institutional knowledge compounds, how you ship software you would stake your name on. All of it, encoded into a system that AI employees can operate.
For 30 years, the instinct was always the same: protect the codebase. The product's source code was the asset. Everything else (processes, decisions, tribal knowledge) lived in people's heads, in meetings, in institutional memory that walked out the door with every departure.
Building a machine-native company inverted that. When AI employees do the knowledge and execution work, every decision, process, standard, and piece of institutional knowledge must be encoded. Written down. Versioned. Enforceable. The company itself becomes source code.
This is what makes the model work. It's what makes it improvable: you iterate on the company the way you iterate on a product. It's what makes it compounding: every lesson gets encoded into the system, not trapped in someone's head. And it's what makes it transferable: the company's capability isn't dependent on any individual's memory or experience.
Here is the key: the artifacts get smarter, not the models. With every failure, the governing artifacts pick up the knowledge and the resolution of that failure so it is only paid for once. The next AI employee who encounters the same situation already has the answer encoded in the system. The models stay the same. The company gets better.
This is not a story about AI replacing humans. I stepped back and redesigned the entire company from first principles around current technology. I am able to do that because I have no legacy baggage. Future startups will have the same advantage.
Humans bring two things that AI cannot provide today: judgment and accountability.
Judgment is the cognitive moat. Humans still make better consequential decisions: what to build, who to build it for, whether this product serves the market or misses it. Every loop in the system has a human at its boundary making the calls that require market sense, taste, and experience. Today, that judgment is what separates a machine-native company from an expensive demo.
Accountability is the durable moat. Even as AI judgment improves (and it will), someone must bear consequences for decisions. Legal frameworks, regulated industries, customer relationships, and social contracts all require a human who is answerable. Accountability doesn't automate away with better models.
One thing I have learned: AI amplifies. It amplifies quick, high-quality decisions the same way it amplifies low-quality ones. The most valuable human asset right now is the judgment to make quick, high-quality decisions that AI can then amplify across the entire organization. Get the decision right and the system multiplies it. Get it wrong and the system multiplies that too.
As AI capability advances, judgment may eventually shift to machines. What remains is accountability. The future of human work in a machine-native company may not be deciding what to do. It may simply be standing behind the decision.
The cost structure of starting a software company has collapsed in stages.
Starting a software company meant raising capital on three fronts: infrastructure (servers, data centers, networking), people (every department required human hires: engineering, sales, marketing, support, finance, operations), and go-to-market (acquiring and reaching customers). All three were expensive. All three required capital before revenue.
Public cloud collapsed infrastructure to a rounding error. AWS, Azure, and Google made it possible to start at zero and scale spend with product-market fit. The playbook shifted: raise for people and go-to-market only. Infrastructure became a variable cost that grew with success, not a fixed cost you raised for upfront.
AI is collapsing people costs the same way cloud collapsed infrastructure. Not just engineering: every function that used to require a human hire. My AI engineering payroll is the API bill, averaging under $1,000 a month in its first seven months, peaking near $2,100. Engineering runs on-demand, not as a steady-state burn.
What's left is audience. If you have a captive audience you understand and can reach, monetization follows. You know what they need, you can build it in days, and you can deliver it directly. If you don't have that audience, that's the one remaining axis that requires either capital or calendar time. Not capital for people or systems. Capital for acquiring the audience itself.
I have seen this show before. During the analog-to-digital transformation there was enormous hype, and then trillions in market value evaporated. In the ashes of that correction, the current giants emerged including Google, Amazon, Facebook, and Netflix. The constraint at the time was bandwidth and distribution of the technology itself. There are no such constraints in this transformation. The models are available via API. The compute is elastic. The infrastructure is built. This will go faster and be more disruptive than anything the software industry has experienced.
It's a rebirth, not a next chapter.
Jason Burns
Lifelong technologist
The last thing the industry needs is another maturity scale, so I resisted making one. But every conversation about this work hits the same wall: the words blur, and none of the labels ask the question that actually separates one business from another. This scale measures the one axis the others do not: who does the knowledge and execution work.
Level 0 · No AI. Humans do the work with conventional tools.
Level 1 · AI-Assisted. Humans do the work. AI drafts, suggests, and accelerates: copilots, chat, coding assistants. Adoption is personal and informal; each person decides how much help to take. This is where most of the industry operates today, and for many businesses it is all that is ever needed.
Level 2 · AI-Embedded. AI stops being personal and becomes organizational. The business formalizes AI into its workflows and systems: purpose-built applications, orchestrated pipelines, sometimes a central AI operations team that builds and runs them. In some companies the copilots fade away entirely because AI is only used through the applications and workflows the business built. In others they coexist. Either way, humans remain involved in the knowledge and execution work, sometimes visibly, sometimes as the shadow human holding a loop up.
Level 3 · Machine-Native. Machines do the knowledge and execution work. Humans hold judgment and accountability at designed boundaries, and nothing else: no shadow humans, because every human touchpoint is deliberate, tracked, and measured. Getting here is not an adoption exercise. The process, the code, the verification, and the organization itself all have to be redesigned around the machines.
Level 4 · Machine-Autonomous. Machines take the judgment too: what to build, for whom, whether it serves the market. No one operates here today, and capability is not the only barrier. Accountability, the world's willingness to accept a machine as the answerable party, may be the one thing that never transfers. If it does not, Level 4 still ends with a human standing behind the decision.
Not every business should be driving beyond Level 1, and the stakes rise with every step. Level 1 is easy: individuals pick up tools and the business barely changes. Level 2 is harder: workflows, systems, and roles get rebuilt, and the disruption is real. Level 3 is a full migration, a redesign of the business from the ground up, with a price measured in more than dollars. When that price pays off is a question I am still testing with my own company.
What I am sure of: a new company gets one chance to choose its starting level, and day one is the easiest and cheapest that choice will ever be. Nothing to unwind, nothing to migrate. There is no right or wrong answer on the scale, only the right one for the business. Choose deliberately and start there. And whatever Level 4 turns out to require, machine-native companies will likely get there first.
Existing companies have a harder job: an honest assessment. Where is the business today? Why do you believe it needs an upgrade? What metrics back that belief? And what will the move truly cost? For a running business, this is swapping a 747 for a space shuttle at 35,000 feet. Not for the faint of heart.
Placing a business on this scale, and pricing the move honestly, is the work I do for a small number of clients. It is described in What We Do.
This work is not finished. It continues until one of two things happens: I hit a true ceiling in the models, or I work through every part of a software company and remove the human from the knowledge and execution work. Every barrier along the way has had the same shape: something a human was quietly doing that the machines could not yet do without one. Each one had to be understood before it could be removed, and each one left behind a concept worth writing down.
Every AI employee needs a place to exist: somewhere to wake up, do their work, and be reached. I chose to build and own that place, a choice that is not right for everyone but pays me back daily. My infrastructure bill runs less than $5 a day, an employee costs nothing when it is not working, and because I own the runtime I can tune it as fast as I learn. This is the story of where my employees live.
Read more →My AI employees are brilliant, tireless, and sincerely wrong at the worst possible moments. One of my engineers once swore his feature worked. He had never actually run it. I cannot stand over thirteen shoulders, so I stopped asking anyone to be trustworthy and built a company where nothing is trusted: every piece of work arrives presumed wrong and has to prove itself before it moves forward. The strange part is what fell out the other side, an audit trail a regulator would envy that nobody sat down to write.
Read more →Every one of my employees starts each morning with total amnesia. Nothing survives in anyone’s head: not yesterday’s decisions, not last month’s lessons, not the reason we do things the way we do. A human company would collapse in a day. Mine gets smarter every week, because the knowledge never lived in the employees. It lives in the artifacts they work from, and the artifacts are what I manage, version, and improve. My employees forget everything, and the company remembers it all.
Read more →Every convention in software exists for a good reason: teams of people building together needed code they could navigate, share, and hand off. The patterns, the frameworks, the abstractions are tools built for those teams, and they served us well for fifty years. But my engineers are machines now, and machines need different things from code than teams of people do. So I started over and asked what those things actually are. I got as far as questioning whether we needed source code at all before my own AI employees pulled me back: they wanted the code kept readable so the human doing oversight could open any file and understand it. What emerged organizes a codebase around what a machine needs to find, understand, change, and verify.
Read more →When no human reads the code, how do you know it is right? My first answer was Craig, my AI review engineer, and I watched him spend his attention the way human reviewers always have: mostly on things that never needed judgment in the first place. Standards, dependencies, contracts, structure. Almost all of it has a deterministic answer. So I split verification in two. Everything with a right answer became a mechanical check that runs in seconds and cannot be argued with. There are 51 of them today, and the number grows with every lesson. What remains for Craig is the part that deserves a reviewer: whether the work is actually good. The discovery was not the checks. It was learning where the line between the two belongs.
Read more →The first loop was almost a toy: AI employees running a LinkedIn content calendar so I did not have to. But watching it run, I noticed the shape of the thing. Standards in, work out, a human at the boundary signing off, and the whole cycle improving itself as it turned. I rebuilt engineering as the same shape, and it held on the hardest function I know. That was the day the org chart stopped being a diagram and became a build order. The toy proved the shape. Engineering proved the company. Four functions to go.
Read more →This company has been recorded since the day it was born. Multiple repositories, versioned from February 11, 2026 onward, hold every design decision, every learning, every improvement, every failure. As of September 27, 2026: 23,649 git commits, 22,165 of them made by AI employees doing their work. 14,428 Slack messages across 1,033 threads, human to AI and AI to AI, 3,095 of them from the one human. There are no direct messages anywhere. Company policy forbids them and the harness cannot send one, so the record is not most of what happened. It is everything. It may be the first company ever documented like the Truman Show.
Running this company keeps handing me discoveries I did not go looking for. A single sentence added to an employee’s identity changed how he investigates. My Chief of Staff described the difference between knowing and searching before I had words for it. Some of what I have observed I still cannot fully explain. Each observation gets a number, a date, and an honest label for how well I can back it up. The registry grows as the company runs.
Read more →Some of what I do every day sounds made up until you watch it. Coaching a model out of a bad habit by taking the blame for its mistake myself. Renaming a tool so an employee stops avoiding it. When a technique is easier to show than to write, I record a few minutes of video and explain it the way I would across a desk. My video blog: hard-fought wins from real failures, a few minutes at a time.
Read more →