Super Intelligence Engineering is the Last Job
Nine days ago, I wrote What Is Applied Super Intelligence Engineering? I described a new job: choose the problem, supply the context, connect the tools, define the check, and let the super intelligence solve it.
Today I removed one word from my title. I am a Super Intelligence Engineer.
The word “applied” made the job sound like a specialty, one career path among many. It is not. The evidence from this year points one way. Super intelligence does the work, including the work of building the next super intelligence. The human part that remains is the same in every field: set the direction, and check the result. Every job is turning into that job. It is the last job.
The hardest test
If any job were going to be fully automated first, it would be the job of building super intelligence. The AI labs have the best models. They have the strongest reason to automate, because each step of automation makes the next model arrive sooner. And they measure their own work more carefully than anyone.
So look at what the labs report.
At Anthropic, more than 80% of the code merged into the codebase was written by Claude as of May 2026. Before Claude Code launched in February 2025, that share was in the low single digits. In a later measurement, Anthropic reports that Claude “leads” 26% of its AI research and development work, up from under 1% in February 2026. About 30,000 agents do research and engineering work there at any one time.
At OpenAI, the goal was an automated research intern by September 2026. On September 6, OpenAI said it reached that goal. The next goal is “an automated AI researcher by March of 2028.” By mid-August, OpenAI researchers used 3.1 agent-workdays for each human workday.
Dario Amodei wrote in January that the loop of AI building AI “has already started, and will accelerate rapidly in the coming months and years.”
Now read what the same reports say about the humans.
Anthropic: “Claude is not operating fully autonomously for any measured subset of AI R&D work.” A human engineer reviews each change that Claude leads and decides whether to ship it. Anthropic also wrote: “Direction-setting was the only meaningful role a human played.” It expects its people to move “most of our effort towards oversight, validation, and verification.”
OpenAI defines its research intern as “a system that can carry out well-defined research tasks under human direction.” More than half of the successful 4–8 hour agent tasks still needed at least one human intervention. Jakub Pachocki, OpenAI’s chief scientist, told MIT Technology Review: “you still want people in charge and setting the goals.”
The labs are automating the frontier itself. What they keep for humans is direction, oversight and verification. That is super intelligence engineering. If it survives at the labs, where automation goes fastest, it survives everywhere.
Every job is turning into this job
The same shape now spreads to all office work.
In 2025, Microsoft gave the new worker a name: the “agent boss,” “someone who builds, delegates to, and manages agents to amplify their impact.” Its 2026 survey of 20,000 AI users in 10 countries found that workers are moving from doing tasks to designing workflows and handing work to agents. In Microsoft’s words: “The people moving the fastest are those that decided where to delegate and what to own.”
Andy Jassy told Amazon’s employees that “there will be billions of these agents, across every company and in every imaginable field.” At a CFO dinner this year, MIT’s Michael Schrage said “CEOs will be judged as much for their agents as for their hires.”
The job market follows. The forward deployed engineer puts agents to work inside a customer’s business. In India, Indeed postings for that role in June 2026 were up 1,807% from a year earlier. Small teams now do what large companies did. Chai reports $100 million in annual recurring revenue with a team of about 15.
The US Bureau of Labor Statistics projects that customer service representatives will lose 153,700 jobs by 2034, a fall of 5.5%. In the same projection, software developers gain 267,700 jobs, up 15.8%, and data scientists grow 33.5%. The jobs that do a task shrink. The jobs that build and direct the systems grow.
The door closes first for the young
The first people to feel this are the ones trying to enter the old jobs.
Stanford’s Digital Economy Lab finds that employment of workers aged 22 to 25 in AI-exposed jobs “now stands 19% below where it would be” if it had kept pace with less-exposed work. A year ago, that gap was 13%. Experienced workers show no similar gap. The cause is mainly that companies hire fewer young people, not that they fire more. Anthropic’s own labor study finds that the job-finding rate for young workers in the most exposed jobs fell about 14% from 2022.
Through September, AI was cited in 120,136 announced job cuts in 2026. It is the leading reason this year. The IMF studied 700 million job ads and found that hiring demand for programmers fell about 10% relative to less-exposed jobs. Undergraduate enrollment in computer and information sciences at four-year schools fell 8.4% in spring 2026.
Students see that the entry-level version of the old job is going away. They are right. The entry-level version of the new job is open.
What the skeptics get right
The skeptics have data too, and it deserves a fair reading.
The Yale Budget Lab says the labor market “largely reflects stability, not major disruption at an economy-wide level.” In a survey of about 6,000 executives, more than 90% report no effect of AI on employment in the past three years. Total announced job cuts in 2026 are down 39% from 2025. Andy Challenger warns that “Naming AI in a layoff announcement can win over investors,” and that AI “is shifting the labor market, it is not dismantling it.”
Some companies cut people for AI and then reversed. In a Robert Half survey, 32% of US hiring managers had cut a role mainly because of AI and then brought the work back. Klarna started to hire human support staff again. Duolingo dropped its plan to grade employees on AI use. Its CEO said: “At the end, we backtracked.”
I do not think this data goes against the thesis. It supports it.
Look at what failed. A company removed the people and kept no one to set the direction or check the output. The agent did a task without a goal it could verify and without a human who owned the result. Quality fell, and the work came back. That is not evidence that humans must do the task. It is evidence that someone must do the super intelligence engineering. The companies that cut the engineer along with the worker found out what the last job is.
The total numbers are stable for now because the change is in the shape of work, not yet in its amount. Fewer young people enter the old jobs. More people manage agents. Total employment moves slowly while the job inside it changes completely.
Why it is the last job
Every other job is a function that a machine can learn. I made that argument in All Human Labor Is Better Done by Machines Now. So why would directing the machine be different?
Part of it will not be different. Anthropic writes that “research taste” might be “just another AI capability that AI systems fail at for a time, then get good at.” I expect that. The machine will get better at choosing the next step, at writing its own checks, and at noticing its own mistakes. The share of the work that needs a human will keep shrinking.
But one part cannot be handed over without losing the point of the work. The work exists to serve people. Someone has to say what we want: which disease to cure first, which product to build, what risk is acceptable, and when the result is good enough to ship. Someone has to own the outcome. A super intelligence can propose a goal. It cannot want the goal for us. When the machine does every task, the human who chooses the goals and accepts the results is the last link in the chain, and the most important one.
That is the job I described in The Only Thing Left To Do Is Unblock The Super Intelligence and in We Are All Managers of Super Intelligence Now. It is the job I do every day with a drone army of agents. It does not need the word “applied.” It is simply engineering, with super intelligence as the material.
The public already knows
Most people already feel this change. In a Pew survey from June, 71% of US adults said AI will lead to fewer jobs over the next 20 years, up from 64% in 2024. Across 37 countries, Pew found that in 34 of them more people expect fewer jobs than more.
The government has started to use the right words. Executive Order 14434 tells federal agencies to say “Super Intelligence” instead of “Artificial Intelligence.” But the order has no plan for work. Neither does anyone else yet.
You do not need to wait for a plan. Take the last job now.
- Stop doing the task. If an agent can do it, give it to the agent.
- Write the goal. State what done means, in a form a machine can check.
- Give it what it needs. Context, tools, credentials, compute and money.
- Check the result. Know your field well enough to catch a fluent mistake.
- Own the outcome. Decide what ships, and answer for it.
- Run many at once. Your limit is your attention, not your hands.
Every other job is becoming this one. The sooner you start, the more of the future you get to direct.
Super intelligence engineering is the last job. It is also the best job humans have ever had.