AI Coding Agent

Software consists of text: instructions that a computer executes line by line. This text is normally written by humans. An AI coding agent is a program that takes on this work itself. You describe to it in plain language what is missing or broken. It then searches the project for the relevant files, modifies them, runs the program, and checks whether it works. If not, it tries again in a different way. It is precisely this independent continuation of work over many steps that distinguishes it from a simple assistant that only suggests individual lines.

What changes for developers and companies

Programming consists to a considerable extent of routine work. Old libraries need to be migrated to new versions. Error messages need to be traced. Tests need to be written, even though nobody likes writing them. Agents can take over parts of exactly this work. This shifts the developer’s role: they describe and review more, and type less.

For companies, this is primarily a matter of cost. Developer hours are expensive, computing time is comparatively cheap. Large software companies now report that a double-digit percentage of their code is machine-generated. Such figures should be read with caution, since it remains unclear exactly what is being counted. Nevertheless, the trend is unmistakable, and it has been driving up the valuations of relevant vendors since 2024.

There is a counterargument. Code that nobody has properly understood becomes a problem later on. Experts speak of technical debt: you save time today and pay it back later with interest. Studies also suggest that the perceived time savings are often greater than the measured ones. Reviewing and reworking also costs hours.

The cycle of plan, tool, and check

At its core is a language model, i.e. an AI system trained to generate meaningful text continuations. On its own, it could only formulate suggestions. It only becomes an agent through tools. It is allowed to open files, search the project for keywords, execute commands, and run tests. The model decides at each step which tool makes sense to use next.

The process is a loop. The agent thinks, carries out an action, reads the result, and thinks again. An error message from a test run is the most important piece of information here. It replaces the feedback that a human would otherwise have to give. That is why agents work considerably better in projects with good automated tests than without them.

The limitation lies in overview. A model can only take a limited amount of text into account at once, the so-called context window. A large project with millions of lines does not fit into it. The agent therefore has to search purposefully and select what it reads. If it misses an important spot, it builds a solution that breaks something elsewhere. For this reason, most agents run in a sandboxed environment and first submit their changes to a human for approval.

From Copilot to the ticketing system

Well-known examples include GitHub Copilot in its agent mode, Claude Code, OpenAI Codex, Cursor, and Devin. Some run as a program in the terminal, i.e. in the computer’s plain text input. Others are built into a development environment. A third group is connected directly to a team’s task system: you create a ticket, the agent processes it, and delivers a proposed change.

In news reports, coding agents often come up in connection with benchmarks, i.e. standardized tests. The best known is called SWE-bench and uses real bug reports from open-source software projects. When a vendor announces that its model solves a certain percentage of these tasks, an agent is usually involved.

A common misconception is that such agents make programming knowledge unnecessary. Rather, they shift where knowledge is needed. Anyone who cannot judge whether a change is clean and safe will notice mistakes too late. The technical term for uncritically adopted AI code is now vibe coding. For small projects, this is often sufficient. For software on which money or safety depends, it is not.