Coding Agent

A coding agent is a program that uses artificial intelligence to write and modify software on its own. You give it a task in plain language, for example: “Fix the login bug.” The agent then independently searches for the relevant files, changes them, tries out the result, and corrects itself if something doesn’t work. The important distinction here is from a simple suggestion tool. That kind of tool only completes the sentence the human is currently typing. An agent works toward a goal over many steps, without anyone triggering each individual step.

Why it matters

Programming consists largely of routine work. Searching for bugs, updating libraries, keeping tests up to date, adjusting documentation. This is precisely the kind of work agents can now often take over. That changes how much a developer can accomplish in a day.

For companies, this is a cost issue, and therefore a stock market issue. Companies like Microsoft, Google, and Anthropic sell such tools as subscriptions. Analysts measure their success by how many tasks an agent solves without assistance. At the same time, there is a serious debate about the downside: anyone who hasn’t written code themselves understands it less well. And an agent can produce convincing-sounding nonsense. Verification remains a human job, especially for software that moves money or protects data.

How it works

At its core is a language model. This is an AI system that has learned to predict the next meaningful sequence of characters from vast amounts of text — for code just as for prose. On its own, this model could only output text. It only becomes an agent through tools.

Such tools are commands the model is allowed to trigger: opening a file, searching the project for a term, replacing a line, running the tests. After each command, the model receives the result back, such as an error message. From this, it derives the next step. This loop of planning, acting, and observing runs until the task appears to be done.

A comparison works well here: a suggestion tool is like a colleague looking over your shoulder and whispering words. An agent is like an intern who is given an assignment and comes back with a finished proposal. With the intern, you examine the result closely before it goes into use. That’s exactly why there is a fixed process: the agent submits its changes as a proposal, a human reads it and approves or rejects it. Many agents also work in an isolated environment so they can’t destroy anything on the real system.

Where you encounter the term

Most commonly in product names. GitHub Copilot, Claude Code, Cursor, Google Jules, and OpenAI Codex are well-known examples. Some run in the editor on your own machine, others in the terminal, and others directly on the platform where the source code is hosted.

In business news, coding agents come up in two contexts. First, in quarterly earnings, because subscriptions with AI features noticeably generate revenue. Second, in the question of how entry-level jobs in the software industry are developing. Anyone who programs themselves encounters the term at the latest in school projects or computer science class. Working with it is a skill in its own right: describing the task precisely, critically reading the result, never adopting anything blindly.