Programming with AI: describing instead of typing

Programming with AI: the new everyday

A developer starts the task with a text box. Three sentences: what the software should be able to do in future, which rules apply, what stays untouched. Then an AI agent takes over, that is, a program that works through a task independently in several steps. It finds the relevant places in the code, changes several files, runs the tests and submits a proposed change. The developer reads, discards, corrects, sends it back. This is how the trade magazine iX described the agent mode of a widely used development tool in September 2026.

The work thus shifts from writing line by line to describing, reviewing and steering. At Stack Overflow, with more than 49,000 responses from 177 countries, 84 per cent of those surveyed in 2025 were using AI tools or planning to. Google's DORA report counts 90 per cent users among just under 5,000 professionals, who spend a median of two hours a day working with AI. Clients know a different kind of programming from quotes and timesheets.

From suggestion to agent

The first stage was autocompletion: the tool suggests the next line, the human accepts it or keeps typing. GitHub, provider of such an assistant, had 95 developers program a small web server in its own 2022 experiment. The group with the assistant finished 55 per cent faster. The second stage is chat: the developer describes a problem in everyday language, gets finished code back and builds it in.

The third stage is agents. An agent is given a whole task. It reads into the existing code, changes several files, runs tests and feeds the results into the next step. On GitHub, a platform with more than 180 million developers, such an agent produced more than a million proposed changes between May and September 2025. Each one went to a human for approval. The majority still work without agents: in the Stack Overflow survey, 14 per cent used them daily, 38 per cent had no plans to.

Vibe coding and its limits

There has been a word for describing instead of typing since February 2025. The AI researcher Andrej Karpathy called it "vibe coding": giving in fully to the vibes and forgetting that the code even exists. He meant it for throwaway weekend projects. The Collins dictionary nonetheless made the term its word of the year in November 2025. It defines it as programming with AI that is instructed in everyday language.

Professional developers keep their distance. In the Stack Overflow survey, 72 per cent said vibe coding was not part of their professional work, a further 5 per cent denied it emphatically. Karpathy himself now prefers to speak of "agentic engineering", and that describes the way companies work more accurately: the human sets the goal, the agent carries it out, the human reviews it and remains responsible.

Describe, review, steer

Describing means more than typing in a wish. The agent needs a goal, rules and context: how the existing system is built, what it must not do, how success is measured. Whatever is missing from the description, the AI fills in by probability, and the result looks finished. That is exactly where the biggest frustration lies: 66 per cent of those surveyed by Stack Overflow cite AI solutions that are almost right, but not quite. 45 per cent find debugging AI code more time-consuming.

Reviewing therefore becomes the main work. Only 3 per cent have a high degree of trust in the accuracy of the tools, 46 per cent distrust it. Those who do not trust ask colleagues: three quarters then turn to a human. Finally, steering means deciding what the machine gets in the first place. According to a JetBrains survey of 24,500 developers, they are happy to hand over routine work, such as scaffolding code and documentation, but keep debugging and the design of the program logic. Even the programming languages are adapting: TypeScript, which catches many errors before the program even starts, became the most used language on GitHub in 2025, according to the platform partly because these checks catch AI errors.

Faster, but not everywhere

Whether all this is actually faster is less clear than the advertising promises. In 2025, the research institute METR had sixteen experienced developers solve 246 real tasks in their own, publicly accessible software projects. By lot, they worked sometimes with and sometimes without AI. With AI they took 19 per cent longer. Afterwards they believed they had been 20 per cent faster. The authors stress that their participants know their projects well, apply high standards and do not represent all development work. Later measurements at the end of 2025 were closer to zero, with considerable uncertainty, and METR conceded that the experimental set-up tends to underestimate the benefit.

At company level, the DORA report sees more throughput and less stability: teams ship more, and things go wrong more often. The authors call AI an amplifier. Teams with automated tests, version control and fast feedback get better. Teams without these foundations ship their weaknesses faster.

What changes for clients

Anyone having software built, in their own IT department or by a service provider, notices the shift in three places. First, the description becomes the most important raw material. Gaps in the brief are filled by the machine according to probability, and that shows up later than a query would. Second, the effort moves: fewer hours for writing, more for reviewing and testing. A quote that advertises only faster writing leaves out the second half of the bill. Third, the benefit depends on the team's basic equipment. Without tests and orderly processes, the software does not get better, only finished faster.

In German companies, this is just beginning. According to the 2026 Bitkom survey, 57 per cent of companies with 20 or more employees use AI. In the IT department it is only 9 per cent, up from 2 per cent the year before. The jump within one year shows the direction. Anyone commissioning a software project today will, with some probability, get code that was described rather than typed. The question of how it was reviewed therefore belongs in the project conversation just as much as the question of the hourly rate.