Researchers mapped how coding AI agents actually work — revealing 13 different architectures where everyone thought there was one

Computer scientists analyzed 13 open-source AI coding agents and found they're built completely differently from each other — different control loops, different numbers of tools, different ways of managing memory and context. This means when someone claims an AI agent can write code, you have no idea what it's actually doing, because there's no standard architecture yet.

For years, AI research has described coding agents by their capabilities: they can use tools, they can plan, they can test their own code. But capabilities don't tell you how the system actually works. This paper opens the black box and shows the scaffolding varies so widely that two agents claiming the same capability might be built on completely different principles. That matters because it means researchers studying why agents fail, or companies building new ones, are working without a shared language. The taxonomy itself becomes useful the moment people start using it — right now, there's nothing to use.