Minimalist Coding Harness

Re: I Hated Every Coding Agent, So I Built My Own — Mario Zechner (Pi) - YouTube

I have customized Emacs and VS Code to suit the way that I like to use computers. Zechner’s talk argues for a world where there is a basic coding agent (Pi) and extensions that are built on top of this agent. Users choose how to compose the agent together with extensions, depending on the way that they like to work. The coding agent does not make the decisions that should be left to users. In my use of Claude Code and OpenCode at work¹, I have noticed that Claude Code has a tendency to “commit” each small change immediately after making it. This keeps commits small, but makes it hard for me to distinguish between AI-written code and code that was already reviewed and committed by me. I don’t like this. I would like to take the “commit” step out of the agent loop (if it is actually defined in there). OpenCode does not do this: It makes changes to files on disk, and leaves everything else to me: This is the way that I like to work.

Both approaches are valid; both will have its userbase; users should be able to customize a single stable coding harness to their needs. They should not have to “switch” to a new harness to avoid features they don’t like or to get features they want to use.

As Zechner says in this talk, no one knows what the ultimate form of a coding agent will look like. My bet is that it will look similar to OpenCode. The TUI separates the agent’s work which happens in the terminal, from the reviewer’s work which happens in an IDE. This separation also creates a distinct step between “write” and “review”, reducing the context switching required. It is an exciting time for tooling enthusiasts!