Codex Agents

Software consists of text files full of instructions, known as source code. Until now, humans typed this text line by line. Codex agents take over part of this work themselves. You describe to them in ordinary sentences what the program should do, for example: “Add a login page and make sure that incorrect passwords produce an error message.” The agent then reads the existing files, changes them, tries out the result, and in the end presents a finished proposed change. The key word here is agent: the system doesn’t just respond, it acts independently across multiple steps.

From suggestion to independent task

The first generation of programming aids worked like a very good autocomplete. You typed the beginning of a line, and the tool suggested the continuation. The human remained constantly in control and had to judge every suggestion immediately. Codex agents shift this boundary significantly. They receive a task that takes ten minutes or an hour of work, and only report back once they have a result.

Economically, this is the reason why such systems are currently receiving so much attention. Programmers are among the most expensive skilled workers in many companies. Yet large parts of their work are routine: writing tests, swapping out old libraries, hunting for bugs in someone else’s code. That is exactly where agents are strongest. Companies such as Microsoft, Google, and OpenAI report that a noticeable share of their new code is already being generated by machine.

At the same time, a new problem arises: review work. Anyone who hands off an hour of work to a machine must subsequently spend an hour checking the result. Errors here are often inconspicuous, because the code looks clean and yet still computes something wrong. Experience does not become superfluous as a result, but shifts from writing to judging.

The cycle of changing, running, and refining

At its core is a language model, that is, an AI system that has learned from enormous amounts of text to produce plausible continuations. It was trained, among other things, on publicly available source code. A language model alone, however, can only output text. It only becomes an agent through tools it is allowed to operate itself: opening files, saving files, running commands on the command line, starting tests.

The process is a loop. The model considers what step makes sense next, carries it out, and reads the result. If a test fails, the agent sees the error message and attempts a correction. This loop repeats until the task appears to be done or a limit is reached. The feedback from actual test runs is the most important difference from a pure chatbot here: the agent doesn’t just guess, it verifies.

To ensure nothing unintended happens in the process, the work usually runs in a sandbox. This is an isolated environment in which the agent is allowed to change files but has no access to the rest of the system. In the end, there is a proposed change that a human approves or rejects. A typical misconception is the assumption that the agent understands the project the way a colleague would. It works with what it sees in the context window, that is, the limited snippet of text it currently has in front of it.

Where such agents are used today

Codex agents are most visible in development environments and on platforms like GitHub, where software is managed collaboratively. There, an agent is assigned a reported bug description and delivers a proposed fix in return. Competing products are called Claude Code, Gemini Code Assist, or GitHub Copilot. The terms blur together in everyday use, but they usually refer to the same idea.

In business news, Codex agents mainly come up in two contexts. First, in connection with providers' revenue figures, since programming is currently the most profitable application of AI by far. Second, in connection with the question of jobs, especially for entry-level workers, whose typical tasks can partly be automated.

You can try this out even without any training. Many tools generate a small website or a game from a description and display every intermediate step. For serious projects, however, it still holds true that without understanding the code, you cannot judge whether the result is actually correct.