This is still the header! Main site

On How the Idea of Projects / Workspaces is Bad

2026/11/22

Many pieces of software have them. Visual Studio Code opens up in a certain root directory. Eclipse has a "workspace". More recently, pretty much all the coding agents have this notion: if you launch e.g. Claude Code in a directory, that's now a "project"; you can resume sessions from here, memories are saved specifically for this directory, etc.

This feels... somewhat normal? It's worth thinking about why we're doing this though. Apart apart from "we have always done this".

After all, different projects need different things in context, right? Different settings, ways of compiling things, etc. Plus the permissions: should the agent have permissions to touch any random other directory? And "projects" and "directories" are kinda the same thing...?

As a contrasting model though... take Emacs. The typical Emacs user will log in, launch an Emacs instance, and keep it around for as long as possible. (I've had some going for more than a year.) It's an extension of your operating system, basically: you can be doing multiple things at once, work in any directory, with none of them especially privileged. There is no artifical boundaries: if the topic you're working on spans multiple directories, you don't have to hack your way around it.

Which is not to say that per-directory configuration is pointless. It is very useful for your editor to know how to launch the current piece of code you're editing; "which directory / git repo is this in" is a useful clue! The main point though is... you wouldn't want to make your actual editor have separate instances per "top level project directory" for the same reason why you wouldn't want to launch a new VM for each C++ file you edit: it's useful sometimes, but it's not always the right division.

Coding agent harnesses seem to have the same problem. Sometimes, a project fits in a single directory; sometimes, it doesn't. You can't even always predict what's going to happen... e.g. you start with a single, well-contained C++ project, but then you have to set up CI for it & that's a different git repo, with all the configs.

Yes, you can open up the config repo separately. On the other hand... why would you? It's just yet another thing to keep track in your head. Is this the CI agent window or the actual C++ project agent window? Which set of settings does it have?

We're doing "projects" even though we already have good solutions for these problems! AGENTS.md files loaded into context when you touch a certain directory solves one side of it. You already are getting the right context depending what you're working on...? Meanwhile, the "permissions" problem is also solved, by "auto" classifiers: agents already are reasonably cautious most of the time, classifiers catch the rest, and if they don't not, project structure is not the thing that is going to save you from the DROP DATABASE equivalent.

And with all this in place, you can just not care about projects or directories. You can talk to The Computer that has been set up to do whatever you want, including finding the code / settings / etc that it needs to accomplish this.

In fact, even today's harnesses, if you just launch your agents in the same directory all the tme & point them to some others, they're perfectly capable of ignoring the part where "projects" were supposed to be a thing. There is still some extra friction, but it's the mindset that counts.