Wiki · Several agents, one repository

CLI coding agent

A coding agent that runs as a command-line process on your machine (Claude Code, Codex, OpenCode, Gemini CLI) rather than inside an editor or behind a web app. It reads and writes your real files with your real permissions, which is both the appeal and the risk.

The category emerged because the terminal already had everything an agent needs and an editor did not: a working directory, a shell, your credentials, your tools, and no sandbox in the way.

What distinguishes it from an editor assistant

An editor assistant is scoped to the buffer and asks permission to touch anything else. A CLI agent starts from the opposite premise. It has a shell, so it can read any file, run any command, install anything, and commit.

That is why it can do end-to-end work and why it needs supervision. The guardrails are not technical restrictions inside the tool; they are the ones you put around it: which directory it starts in, whether that directory is a git worktree per agent or your actual checkout, and whether anything it does reaches a remote without a human.

The practical consequences

It is interactive. It redraws, prompts and takes single keypresses, which is why driving one programmatically means a pty rather than a pipe.

Its billing is tied to the binary. A subscription authorises the official CLI, not the API. See subscription vs api billing.

It has its own tool set. Whatever the installed CLI can do, it can do; a wrapper that re-implements a subset gets a different, smaller agent.

Why several at once is the hard part

One is easy to watch. Three is where it breaks down: they finish at different times, two of them are waiting on a question you have not noticed, and all three are editing the same repository. Everything difficult about this category is downstream of that.

See also

What points here