Skip to content
Important: We're now taking requests for PHP2Shell emergency response —Learn more →
SolanaLink LogoSolanaLink
AI Agents Are a Lever for Programmers, Not a Replacement

AI Agents Are a Lever for Programmers, Not a Replacement

Image generated by xAI Grok
Development
13 min read
0
0 comments
14 views

AI agents are making the programmer's lever longer than ever, but the fulcrum is still human. Using recent data from the US and Japan, this essay looks at what is really changing in capability, productivity and jobs, and at what grows more valuable as code gets cheaper.

Give Me a Place to Stand

"Give me a lever long enough and a place to stand, and I will move the world." The line is attributed to Archimedes, more than two thousand years ago.

I have been thinking about it a lot lately. My screen usually holds a dozen or more windows, and in each one a Claude Code agent is at work. One is refactoring. One is filling in tests. One has paused and is waiting for me to decide. I have built software for more than twenty years, and never before have I been able to move so many things forward at once, by myself.

But the voices outside are not calm. Headlines about AI replacing programmers arrive one after another. In Stack Overflow's 2026 survey, 17% of developers said they deliberately keep their distance from AI, worried about losing their skills, or about training the very thing that will replace them. A power we have never had before, and an unease that is hard to name. That may be what programmers share right now.

So I went looking for data, from the United States and from Japan, as recent and as close to the source as I could find. By the time I finished, my answer had become clearer, not murkier. AI agents are a lever for programmers, not a replacement. The lever is getting longer; that much is true. But a lever cannot push on its own. It needs a fulcrum. The fulcrum is us.

The Lever Is Already Long

First, how long has this lever become?

In 2025, an engineer at Rakuten handed Claude Code a complex open-source refactoring project that spanned several programming languages. It worked on its own for seven hours straight. The engineer said afterward that he did not write a single line of code in those seven hours.

A year later, seven hours looked small. In June 2026, Anthropic released Claude Fable 5. Stripe used it to migrate a 50-million-line Ruby codebase in a single day, work that would have taken a team more than two months by hand. Mythos 5, the same model offered only to a small group of partners, worked largely on its own for more than a week on a genomics research project. With Opus 5.5, released in September, one early tester finished a 680,000-line code migration in less than a day. Another used it to audit and fix a 200,000-line codebase in under three hours; the previous model had needed more than twenty.

And how does it compare with senior engineers? Every set a test: rebuild a real collaboration system that kept crashing in production. Two human senior engineers scored 89 and 96. Claude Fable 5, given follow-up instructions by a person, scored 91, landing right between them.

Even the rulers built to measure AI are running short. The research group METR measures capability with one question: how long a task, in human expert time, can an AI complete on its own half the time? Since 2023, that length has doubled about every 4.3 months. In March 2026, METR measured at least 16 hours, and then admitted plainly that beyond that point its task suite holds too few long tasks to measure with any accuracy.

All of this is true. If the story ended here, the fear would be justified.

But the Fulcrum Is Human

The story has another half.

Two Harvard economics PhD candidates, Fiona Chen and James Stratton, followed the work records of more than 720,000 employees at 655 companies from January 2021 through August 2026. When these companies added coding agents on top of AI coding assistants, lines of code rose a further 44%, commits 36%, and pull requests 36%. Yet the tasks actually completed rose only 14%, a result that is barely significant, and the paper's estimates rule out any output gain above 31%.

Where did all that extra code go? It was waiting in line. Getting code from review submission to approval now takes 56% longer on average: the wait for a reviewer grew by 60%, and the review itself by 59%. When Faros AI analyzed more than 10,000 developers, it saw the same picture: 98% more merged pull requests, 91% longer reviews, and no faster delivery for the organization as a whole.

There is a subtler finding, too. In 2025, METR ran a randomized controlled trial. Experienced open-source developers using AI on real tasks were actually 19% slower, yet afterward they believed they had been 20% faster. Our feelings can deceive us, and they do it sincerely.

On Snorkel's set of 100 senior-level engineering tasks, today's best models still fall short of the correctness and taste expected of a senior engineer on more than 60% of them.

In my own one-person company, I hit the same wall. With a dozen agents running at once, the bottleneck is no longer writing code. It is how fast I can read the results and make decisions.

Code has become cheap. What has become expensive is judgment. Is this code right? Should this work be done at all? Can we live with this risk? You can keep adding length to a lever. But every time you push, the force lands, in the end, on human judgment.

Two Shores, Two Kinds of Weather

Now, jobs. On the two shores of the Pacific, the weather is very different.

In the United States, the entry door has narrowed. Using ADP payroll data, Stanford's Digital Economy Lab found that in the occupations most exposed to AI, including software development, employment for 22- to 25-year-olds is about 19% below where it would be had it kept pace with peers in less-exposed jobs. A year earlier, the gap was 15%. Most of the change comes not from layoffs but from fewer new hires. Data from the Federal Reserve Bank of New York show that in March 2026, the unemployment rate for college graduates aged 22 to 27 was 5.6%, above the 4.2% for all workers, with engineering and computer science standing out. For young people just leaving school, these numbers mean more applications that never get an answer.

Yet the same Stanford study also says this: there is no sign of widespread displacement across the economy, and experienced workers show no similar gap. In the Bureau of Labor Statistics' latest ten-year projection, software developer jobs are still expected to grow 10%, with about 95,000 openings a year.

In Japan, the weather looks different. Of the university students who graduated in March 2026 and looked for work, 98.0% found jobs. When the Recruit Works Institute surveyed 186 companies with 1,000 or more employees, 68.8% said AI would not change how many new graduates they hire. Data from Levtech show 9.1 IT job openings for every engineer looking to change jobs.

Japan's change is hidden inside the mix of roles. Job postings for software testers have shrunk to 0.17 times their peak of two years ago, while consultant roles have 31.6 openings per applicant and project managers 23.7. Levtech reads this as companies in the IT job-change market shifting their attention from "young and inexperienced" to "specialists ready to contribute on day one." As for new graduates, the Daiichi Life Research Institute explains that Japanese companies hire new graduates in order to train them, so even as AI takes over entry-level tasks, the seats for new graduates do not simply disappear.

The weather differs, but the wind blows the same way. Repetitive work that can be done by following the manual is going out with the tide. Work that calls for judgment, design, and responsibility is coming in.

Who Holds the Lever

So who is holding this ever-longer lever?

More and more often, a single person. In Stack Overflow's 2026 survey, the share of developers working in one-person organizations rose from 4% to 10%. Data from Carta, the equity-management platform used by many U.S. startups, show the share of new companies with a solo founder rising from 23.7% in 2019 to 36.3% in the first half of 2025. Venture capital still prefers teams, though: solo-founded companies started in 2024 received only 14.7% of that year's priced equity funding. Israeli programmer Maor Shlomo built Base44, an AI app-building tool, without any outside funding, and six months later sold it to Wix for $80 million in cash. By then, to be fair, he had eight employees.

Large companies are holding it too, and what they are choosing is to amplify people, not to cut them. After adopting Claude Code, Rakuten shortened the time to release new features from 24 working days to 5. Hitachi has announced that it will bring Claude and other advanced AI into every business process across its group of roughly 290,000 people, and train 100,000 of them as AI professionals.

As writing code got cheaper, the number of people making software went up. GitHub passed 100 million developers in 2023; by September 2026 it had more than 225 million, including about 38.3 million in the United States and 7.15 million in Japan. This September, agents wrote the first draft of more than 40% of pull requests, up from 17% just six months earlier. Software is reaching further, too: public repositories serving mining, construction, healthcare, and agriculture have each grown several times over since 2022.

Economics has an old story about this. In 1865, William Stanley Jevons noticed that the more efficiently steam engines used coal, the more coal Britain burned. Efficiency lowered the cost, and the lower cost released demand. Whether software will follow the same path, only time will tell. But the numbers in front of us already show the first signs.

I am a small footnote to this lever myself. My company is a company of one. Each project has its own development agent. A "project manager" agent gathers their progress and tells me what needs my decision. A "product manager" agent writes the requirements documents. Education, vacation-rental management, travel matching, community e-commerce: products like these once needed a team. Today, what they need is my judgment.

What Is Getting Expensive

If code is getting cheaper, what is getting more expensive? The data agree to a surprising degree.

Verification. Pull requests are piling up, and reviews keep stretching out. In Stack Overflow's survey, nearly half of developers said they trust AI only when its results are easy to check. In 2025, broken access control became the most common alert raised by CodeQL, GitHub's code-scanning tool, flagged in more than 150,000 repositories, and GitHub believes part of the cause is AI-generated scaffolding that skips critical permission checks. The ability to see what is wrong is becoming scarcer than the ability to write code.

Choosing. In METR's survey, people reported that AI made them about 3 times faster, but only 1.4 to 2 times more valuable. Building fast is no longer hard. Knowing what is worth building is.

Experience. Stanford's research also observed a pattern: the more a job relies on knowledge you can learn from books and documents, the more young people's employment in it is falling; the more it relies on practice, the faster experienced people's employment is growing. The researchers caution that this is a correlation, not proof of cause. Every's test pointed the same way: the widest gap between models and humans lay in remembering the lessons of past incidents, and in having the nerve to truly simplify a system.

And design, management, and responsibility for live systems and real users. Only 20% of developers use AI to deploy software or troubleshoot production problems. Only about a third say AI helps them communicate.

These abilities have one thing in common. They grow inside people, slowly, through each failure and each responsibility we carry.

To Everyone Holding the Lever

If you have been writing code for years, please stop thinking of yourself only as the person who writes the code. Your real value is knowing how a problem should be defined, knowing where things will break, and knowing when to say no. Those are exactly the things agents lack most. Starting today, try being the person who defines the problem, directs the agents, and signs off on the results.

If you have just entered this field, I know the door in the United States has narrowed, and that Japanese companies expect more than they used to. But you are holding tools no generation before you ever had. You don't have to wait for an entry-level job to prove yourself. Use agents to build a product that real people actually use, and let your work speak for you.

And whoever you are, judge your own output by data, not by feel. That experiment already warned us: feelings can deceive.

I believe the future of programmers will be better. Not because AI is weak, but precisely because it is strong, strong enough to let one person do what once took a team. GitHub has shared a story like this. A developer in Nepal built a flood bulletin page on GitHub Pages, and later turned it into a service that relays official disaster alerts. No big company, no big team. Just one developer, and the people that developer wanted to protect.

The lever will keep getting longer. I hope we use the extra strength to make someone else's life a little better.

Stay true to yourself, and be kind. In this era, those words deserve a place in our code more than any technology does.

Sources

Capability

Productivity

Jobs: United States

Jobs: Japan (Japanese-language sources)

Scale and cases

Comments

Comments (0)

#Software Development#Productivity#AI#Japan Market
Back to News & Blog