
I Said the Future Was Vibe Coding. Then Goldman Sachs Pulled Back.
In July 2025, I wrote that the future was vibe coding. AI was making it possible to describe an application in ordinary language, generate working code, and refine it in minutes. I still believe that. I built this website with that approach.
But there is a question I did not spend enough time on: what happens when the application is built inside a company, connects to real systems, and becomes something colleagues rely on?
What Goldman Sachs changed
Goldman Sachs has brought that question into focus. According to Bloomberg's November 2026 report, the bank restricted most bankers and traders from using certain advanced AI coding tools after reviewing productivity and risk. Engineers retained access, as did the employees known as strats, whose work spans trading, quantitative research, and software engineering. This was a restriction on particular coding tools, not an abandonment of AI.
The change is striking because, only months earlier, Goldman's chief information officer, Marco Argenti, had praised nontechnical employees using AI to develop ideas and pass them to engineers for further work, as Bloomberg reported. That is exactly the promise of vibe coding: the people closest to a business problem can help build its solution.
We do not know the details of Goldman's internal assessment. Its decision does not prove that employee-built applications failed, or that other companies should impose the same restrictions. But it does expose a question every company will face: how do you give employees the power to build without leaving the resulting software outside the company's controls?
The work behind a successful demo
In Guanta's story, “Anyone can code, but…”, I explored the work that sits behind a successful demo. Someone must decide where data lives, who can access it, how a change reaches production, and who responds when something breaks. AI can generate excellent code without knowing any of those answers. A prototype can be useful and still become an unmanaged dependency.
There is a bigger opportunity here than creating more internal apps. In “Every Company Needs an AI Operating System”, I argued that much of a company's real operation lives between its official systems: in spreadsheets, emails, approvals, reconciliations, exceptions, and the knowledge employees carry in their heads. AI makes more of that work programmable. The employees doing it every day are often the best people to identify what should change.
What piece is missing?
I think it is a way for those employees to become builders of the business operating system, with engineering and operational support around them.
Imagine a trader who repeatedly reconciles information from several systems. They understand the exceptions, delays, and decisions far better than a developer receiving a brief description of the problem. With an approved AI coding environment and safe test data, they could demonstrate a better workflow. That idea could then move through a clear path: review the data access, check the calculations, integrate with the right systems, assign an owner, test it, and measure whether it improves the work.
The employee contributes the business knowledge. Engineers help turn the idea into dependable software. Security and permissions set real boundaries. The company keeps the result, instead of depending on an application that only its creator understands.
This path matters for adoption, too. Training people to prompt an AI tool is a start, but lasting adoption comes when they can see their knowledge improve a real process. They need room to experiment, a simple way to propose changes, and a credible route from prototype to production. Otherwise, the most enthusiastic employees either stop trying or build useful things that the organization cannot safely support.
Goldman Sachs has chosen to narrow access to certain tools. Other companies may make different choices. The lesson I take from the news is that distributing coding capability is only the first step.
I would now extend the development loop I described last year: describe, generate, refine, review, and operate. If companies build that full loop, they can use the talent they already have to create their business operating system—and give people a reason to participate in building it.
Cover image: AI-assisted adaptation of a photograph by Ajay Suresh, licensed under CC BY 4.0.





