AI coding agents and a rerun of the operating-system wars
Photo source: Rod Long (@rodlong) | Unsplash
Attending the Agentic AI for Science workshop at MPCDF in Garching got me thinking about the platforms we are starting to build our scientific workflows around. The workshop programme brings together hands-on work with agents, examples from scientific software development and materials simulations, and perspectives from industry. For someone interested in free and open-source software, this raises some familiar questions: who controls those platforms, what can we change ourselves, and how much of our work can we take elsewhere?
The software development work discussed at the workshop is an example of agentic engineering, a term used by Simon Willison in his guide, Martin Fowler, and Andrej Karpathy. IBM also uses the term, as does the Linux Foundation's AGNTCon + MCPCon programme.
Coding agents use language models to generate code and tools to execute it, while engineers define the problems, guide the work, and verify the results. What interests me here is the software around the model. An agent harness supplies context, exposes tools to the model, and manages their execution. The flagship GPT, Claude, and Gemini models are proprietary, but the software using them does not have to be (and often isn't, as we will discuss below). We can use the same model through different harnesses, with different licenses, tools, and restrictions. Choosing the model, deciding where it runs, and choosing the harness are separate decisions, even when a product offering bundles them together.
A few harnesses and development environments for programming, scientific workflows, and other tasks illustrate these differences. This is only part of the picture: agents embedded in enterprise applications bring their own integrations and dependencies.
This led me to a somewhat loose comparison: Claude Code reminds me of Windows, Codex CLI of BSD, OpenCode and the like of Linux, and Antigravity of the Macintosh. Each comparison emphasizes something different: compatibility, permissive reuse, independence from suppliers, or the appeal of an integrated experience.
If you have used Claude Code in a terminal, the first part probably sounds wrong already. How does a program that reads files, invokes shell tools, and fits into a Unix workflow resemble Windows? Bear with me. The resemblance is in how people build tools and workflows around the software, rather than how its interface looks.