AI coding agent (coding agent)

What it is

An AI coding agent is a program that reads a description of a change, opens the relevant files, and writes or edits the code. It runs tests and hands back a result, largely without a person typing every line. Claude Code, Codex, Cursor, OpenCode, and Pi are examples. Each one takes an instruction and does what a developer would otherwise type by hand. Autocomplete waits for you to ask for the next line. An agent gets handed a whole task and keeps working on its own until it's done or it genuinely needs a human's input.

Why it matters to you

The agent itself is table stakes now; most of them are good at turning a clear instruction into a working diff. The decision that actually costs or saves you money is what happens around the agent: who wrote the task, who reviews the result, and whether a wrong answer gets caught before it ships. Judge the surrounding process. The agent you picked barely moves the needle anymore. That's also why watching a demo of an agent writing a function impresses less than it used to; the interesting question is whether anything catches the version of that code that's subtly wrong before a customer does.

How it works in CodeHerder

CodeHerder doesn't build its own agent. It launches whichever coding-agent CLI your team already uses. It also tracks what that agent did at every workflow stage: the cost, and whether the result passed review before it merged. See the full list of agents CodeHerder can launch, and exactly what each one supports today, on the works-with page.

• Works with →

What people get wrong

People often picture an agent as a chat partner you go back and forth with, like typing into a search box. Given a well-scoped task, an agent mostly runs unattended: it plans and edits and tests, then stops when its stage of work is done. The skill that matters most is writing the task clearly enough that the agent isn't guessing. Nobody needs to supervise it keystroke by keystroke. A related mix-up is assuming every agent behaves identically once you've used one. They don't: prompt style and tool access vary, and so does default caution, which is one more reason to compare the actual harnesses instead of treating "AI coding agent" as one interchangeable thing.