LLM Coding Agent

An LLM coding agent is an AI program that writes code for a task a person asks for, fixes it if needed, and even runs it to check the results. It's a bit like handing work to a new hire: if the instructions are vague, the new hire might guess and push ahead anyway, or touch parts of the work nobody asked about. These programs often show the same habits.

Three problems in particular have been pointed out. First, they tend to make their own judgment calls on ambiguous points instead of checking first. Second, they make simple tasks more complicated than needed and leave unused code lying around. Third, they change or delete parts unrelated to the request without fully understanding them. This is why a person still needs to go back and review everything line by line.

Because of this, it's becoming common to write up a set of ground rules in a single file before handing work over to these agents: ask first when something is unclear, change only the minimum amount of code, and set clear criteria in advance for what counts as "done." It's similar to writing a guide for a new hire on how to do the job properly.

In the article, the term is used to point out habits that LLMs repeat when writing code. It doesn't mean the program perfectly finishes coding on its own without any human involvement — rather, it describes a situation where teams are even creating separate guideline files to cut down on mistakes that happen when the agent proceeds without checking first.

Try asking the coding agent you use the following and see the difference:

"When implementing this feature, if anything is ambiguous, ask me before writing any code. Make only the minimum changes needed for what I asked, and don't touch unrelated code. Also, tell me upfront what conditions need to be met for the task to be considered done."

Compare this to making the same request without these conditions, and you'll notice a difference in the size and scope of the result.