Running coding agents on Windows

The terminal is the part that breaks

An agent CLI is a program that expects a terminal to talk to: a pseudo-terminal it can drive, resize and read interactively. Windows did not have that by default for most of the years this tooling was written, and the emulation layers that appeared instead — a shell wrapping a shell — fail in specific, annoying ways: no resize, no colour, broken arrow keys, a TUI that renders once and then freezes.

The fix is a real PTY: ConPTY on modern Windows, driven directly rather than through a compatibility shim. It is the difference between an agent you can watch work and one you have to run blind.

The smaller paper cuts

• Paths. Unix shells assume forward slashes and a home directory; Windows has drive letters and two conventions. Quote path arguments.
• Line endings. A repository checked out with CRLF makes every diff an agent produces look larger than it is. Set core.autocrlf deliberately, not by default.
• Shell choice. PowerShell, cmd and Git Bash differ in quoting and in which of the two command separators exists. A launch command written for one breaks in another.
• File watching. Directory watchers on Windows are slower and have limits; a large repository can exhaust them with several agents running.

What a working setup looks like

Real PTY terminals, one git worktree per agent, and a launcher that builds the command for the shell that is actually in use rather than assuming a POSIX one. That is the whole list, and it is why the Windows side is worth building deliberately instead of inheriting.

Everything else — the agent, the model, the review — is the same on Windows as anywhere else. The platform should be invisible by the time you are reading a diff.