AI agents can write the code now. A 1995 Seinfeld episode shows the two things still worth owning.

Google reports that 75% of its new code is now AI-generated and approved by engineers, while Anthropic says more than 80% of the code it merges is written by Claude, according to figures cited in an essay on AI coding agents. The essay argues that coding itself is effectively solved and that the remaining value sits in business knowledge held by owners, illustrated through an open-source agent loop called aloop built on top of coding agents such as Claude Code. The aloop approach has an owner state a one-line mission and budget, after which agents write the spec, work test-first in short sprints, and can raise but not lower an approved baseline without the owner.

The Soup Nazi never wrote a spec in his life. He is exactly who AI agents are for. In the 1995 Seinfeld episode named after him, every rule of his soup line lives in his head: know your order, keep it moving, no small talk, no complaints. George complains that he did not get free bread. No soup for him. Elaine makes small talk at the counter. Banned for a year. It is strict, it is fast, and nobody else could have designed it. None of that is code. It’s business knowledge. It is knowing how the business should run, and it has always belonged to the owner . What is new is that the owner no longer has to translate it for anyone. In 1952, almost nobody believed a machine could write code. “I had a running compiler and nobody would touch it,” Grace Hopper recalled https://cs.yale.edu/homes/tap/Files/hopper-wit.html . “They told me computers could only do arithmetic.” Two years later, IBM promised that FORTRAN “ should virtually eliminate coding and debugging https://softwarepreservation.computerhistory.org/FORTRAN/BackusEtAl-Preliminary%20Report-1954.pdf .” Its 1957 paper was titled, plainly, “ The FORTRAN Automatic Coding System https://softwarepreservation.computerhistory.org/FORTRAN/paper/BackusEtAl-FortranAutomaticCodingSystem-1957.pdf .” On coding, they were early, not wrong. Almost nobody writes machine code today. Compilers, then languages, libraries and frameworks, each moved people one level closer to simply saying what they want. AI is the same step, one level higher. Google says 75% of its new code https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai/ is now AI-generated and approved by engineers. Anthropic says more than 80% of the code it merges https://www.anthropic.com/institute/recursive-self-improvement is written by Claude. How to code is solved, the way it always has been: one layer at a time. Look at how that 1954 report splits the cost of a problem: analysis and programming, coding, debugging, and machine time. FORTRAN went after coding and debugging. Analysis and programming, deciding what the machine should do, stayed with people. Seven decades later, it still does. That knowledge does not live in a spec or a developer’s head. It lives with the person who has the problem. The Soup Nazi knows that bread complaints slow the line and small talk stops it. The old way, he would explain that to an analyst, who would write a spec, which developers would turn into code, losing something at every handoff. Spec-driven development is having a moment: write a better spec, get better code from the agent. For most businesses, that has it backwards. The owner should not have to write the spec. The spec does not disappear. The agent writes it. Here is how that works in aloop https://github.com/adeelahmad/aloop , the open-source loop I have been building on top of coding agents like Claude Code. The Soup Nazi gives it a mission in one line, the fastest soup line in New York, and a budget. The agent does the rest. It researches how the best lines run, turns “fastest” into scripted customer journeys, and measures his line against them today. Even the mission starts as a failing test: if every journey already passes, the bar is too low. The gap between today and the mission becomes the first goal. He approves the mission and the baseline once, with a line he types himself. After that, the agents can raise the bar but never lower it without him. Then the loop builds in short sprints, like a disciplined agile team: one task per agent, strictly test-first. One agent writes a test that must fail first, and for the right reason, like a scripted order that still takes too long. Another makes it pass, with no test switched off. Each works in its own copy of the code and sees only its task, so a cheap model is enough. The checks do not live in the model. The model is probabilistic: it can say “done,” mean it, and be wrong. So everything that decides what gets in runs outside it, in deterministic gates: scripts that return pass or fail. If a gate crashes, nothing gets in. A prompt is a suggestion. An exit code is a decision. This part runs today. Across aloop’s 40 public runs https://github.com/adeelahmad/aloop/blob/master/docs/evidence/README.md on Claude Code, failures included, the median run took 13 minutes and cost 89 cents, on small libraries and a command-line tool. In 37 of the 39 runs that counted, a gate said no at least once. The mission layer on top is newer: it passes an end-to-end test offline, but has not had a live run yet. This next part is what I am building now. Shipping is where the loop starts learning. Picture a camera on the line and the order data flowing back. The agent sees the patterns: the line stalls when someone asks for free bread, and again when someone chats at the counter. George and Elaine, as data points. Once there is enough real usage, those patterns shape the next goals. When a goal is met, the loop returns to the mission and raises the bar. Monitor, analyze, plan, execute, repeat: recursive self-improvement of the shop, not the model. The agent still does not change the shop by itself. Anything that reaches his customers waits for his recorded yes, so it arrives as a proposal: a display above the counter that shows the rules and flashes NO SOUP FOR YOU at anyone who breaks them. He approves. He never has to say it again. It also suggests free bread with every soup. No more stalls. His answer: no bread for you. For owners: no spec, no translator, no jargon. Know your business and say it plainly. For developers: the work moves up again, as it always has. Less writing the code. More building the systems that write it, check it and learn from it, and more time understanding the problems they serve. In the end, the Soup Nazi gives Kramer a spare armoire, and it ends up with Elaine. Inside, she finds his secret recipes. She threatens to make them public, and he closes the shop and leaves the country. Look at what ended him. Not a rival, not better code. His recipes, given away inside an armoire he never checked. When anyone can build the app in an afternoon, a business owns two things that matter. Its recipe: what to build, and why. And its data: how real customers behave, which is what keeps the loop improving. Elaine got the recipes. Add the data from his line, and she would not have needed him at all. Let the machines have the code. Keep the recipe, keep the data, and check the armoire. I build aloop in the open at github.com/adeelahmad/aloop . Its sibling, AgentRC , is a draft open spec for declaring what an agent is and what it is allowed to do. AI agents can write the code now. A 1995 Seinfeld episode shows the two things still worth owning. https://pub.towardsai.net/ai-agents-can-write-the-code-now-a-1995-seinfeld-episode-shows-the-two-things-still-worth-owning-b0e0a1de5ea0 was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.