Eleven Seconds, And It Was Correct

A machine wrote a production authentication system in 11 seconds last week. Rate limiting. Session management. OAuth providers. Database migrations. Test coverage. The senior engineer who looked at it said the job takes three days. Eleven seconds is not the story. Correct is the story. I have spent sixteen years deploying technology in places that punish optimism: rural Nigeria, post-conflict Serbia, Kazan, Fukushima. Every machine I ever shipped arrived with a spreadsheet attached. Fuel logistics. Grid cost. Payback measured in years. This one arrived with a login box and a Monday morning. We have stopped asking whether AI can write software. It writes code at a senior level, first attempt, 70 to 80 percent of the time, with nobody touching it. Documentation, refactoring and code review are commoditised. The date on this is mid-2026, not 2035.

The headline number is 70 to 80 percent. The honest number is 40 to 60. That is how much of total coding time evaporates when a skilled developer stops typing and starts reviewing. You become an architect-auditor. You stop building and start approving, and you sign more than you write. Watch who this helps. Senior engineers get dramatically more powerful. Junior engineers get dramatically more redundant. The middle of the ladder falls out first, and it falls out quietly. No announcement. Just fewer seats in the next quarter's plan. What still fails: autonomous debugging that does not introduce two new bugs for every one it fixes. Multi-file changes that break something subtle three directories away. Greenfield system design from genuinely ambiguous requirements. Those fail for one reason. The loop is not closed. The agent writes the code. It does not run it, watch it in production, learn from telemetry and iterate. That loop is the whole game. It is also temporary.

Here is the inversion. Today you write code and occasionally ask AI for help. By 2028 you write specifications and occasionally inspect the generated code, the way you occasionally read assembly output now. Which is to say: almost never. Three things make it real. Context. When an agent holds your entire codebase in working memory, every file and every historical decision encoded in comments and commits, it stops making the mistakes that come from partial understanding. The capability exists today. Cost is the last bottleneck, and it is falling. Verification. The agent writes, runs the tests, reads the error, fixes it and repeats without handing back. The primitive versions are out there and their recovery rate from multi-step failure sits below 50 percent. Cross roughly 85 percent and one agent clears most bug-fix tickets alone. You review architecture, not implementation. Price. A complex pull request costs 50 cents to 2 dollars to generate. Once delegation is cheaper than context-switching, nobody hires for typing speed. The break-even line is a 30-minute human task, and it is already behind us. What you hire now is judgement.

When any competent operator can clone a product's core in a few days, software stops being a differentiator and becomes an input. Like electricity. What dies. CRUD tools with no network effect. Project management. Basic analytics dashboards. Simple workflow automation. Anything where a better interface was the moat, because a polished interface now generates from a screenshot and one paragraph. The API wrapper is the most exposed of all. Your AI-powered CRM is next year's free plugin on the platform underneath it. What compounds. Data where every user makes the product better for the next one. Certifications that take 18 months and cannot be generated in 18 days. Integrations wired into 200 customer systems, which a model can copy in code and never copy in relationships. Regulatory position. And the brand that makes someone type your address instead of generating their own equivalent. Revenue per employee in software runs at 150,000 to 250,000 dollars today. By 2029 the surviving pattern is 1 to 5 million. The 100-person company becomes a 5-person company with a swarm. Profit migrates from the builders to the trust-holders.

Eleven seconds is not the story. Correct is the story.

By 2031 a production-grade product goes from a written specification to running software in hours. Maintenance, security patching and scaling sit with autonomous agents, and human oversight moves up a level: governance, not code, not architecture. The job title full-stack engineer disappears the way the switchboard operator disappeared. It is replaced by product specifier and system auditor. Then the arithmetic turns strange. Why pay 50,000 dollars a year for a tool when an agent builds your equivalent in four hours, maintains it continuously and bends it to your exact workflow? Build-to-suit replaces buy-to-rent across whole categories. And the thing nobody says loudly enough: the defensible asset was never the code. Code was always the thinnest layer. We could not see it, because code used to be expensive to produce. When something expensive becomes free, the layers around it become the entire business.

I signed a EUR 20 million personal guarantee and watched the company fold on Christmas Day 2008. I lost savings in the 2013 Cyprus bail-in. Both events were called improbable by the people who were paid to price them. So I stopped forecasting and started pricing. In 2008 I restructured a EUR 75 million industrial group, and the only reason it survived is that we stopped guessing which quarter the money would stop and started asking what we could live without. We deployed 210 energy systems across Africa and Asia on that logic. Nobody could name the grid that would fail, so every unit had to run without one. Ask the question now. What do you have that an agent serving your competitor cannot copy this afternoon? If the answer is that our code is better, the probability that sentence still means anything in 2029 is close to zero. If the answer is data that compounds with every user, a certification that took two years, integrations too deep to rip out, a brand built across a decade, you are standing where the trajectory is going. The strategic money right now is not in features. It is in infrastructure, compliance, data and distribution. Code is becoming free. Everything around it is getting expensive. Price accordingly.