Agent loop: The cycle a coding agent runs: read the task, look at the repository, call tools (edit a file, run a command), read the result, and repeat until it decides the task is done or gives up.
Coding agent: A program that takes an instruction in plain language and edits a codebase to satisfy it — reading files, running commands and writing code without a person in the loop for each step.
CLI agent: A coding agent that runs in a terminal rather than in an editor or a browser tab. Claude Code, Codex and Gemini CLI are CLI agents.
TUI agent: A CLI agent that draws a full-screen text interface — panels, scrollback, key bindings — rather than printing lines. A TUI agent needs a real terminal, which is where pseudo-terminal support stops being a detail.
git worktree: A second working copy of the same repository, with its own branch and its own index, sharing one object database. `git worktree add ../task -b task` creates one; work done in it cannot touch your main checkout.
Merge queue: An ordered list of finished branches waiting to be merged, so several parallel tasks are integrated one at a time instead of racing each other.
Diff review: Reading what changed before it lands. With agents it is the whole quality gate: the model’s summary is a claim, and the diff is the evidence.
Exit condition: A concrete definition of done for a task — a test that passes, a command that exits zero, a route that answers 200. Without one, an agent stops when it feels finished.
Pseudo-terminal (PTY): The kernel device that makes a program behave like it is talking to a human at a terminal: interactive input, resizing, colour, signals. Agent TUIs need one; a pipe will not do.
ConPTY: Windows’ pseudo-terminal API, introduced to give Windows the same interactive terminal primitive Unix always had. It is what makes a Windows terminal a terminal rather than an emulation.
Swarm: Several agents working in parallel, each in its own worktree, with one merge queue at the end. The name for the mode where the parallelism is the point.
Local-first: Software that does its work on your machine against your files, and does not require a hosted copy of your project to function. Nix Agents is local-first: the agents run where your repositories already are.
BYOK (bring your own key): Using a provider account you already have rather than one bought through the tool. Nix Agents goes further: it does not ask for a key at all, because the CLIs it runs authenticate themselves.
Context window: How much text an agent can hold at once — instructions, file contents, command output. Worktrees and narrow tasks keep the window filled with relevant code instead of noise.
Tool call: One action an agent takes through a defined interface: read a file, write a file, run a command, search the repository. The loop is made of tool calls.
MCP (Model Context Protocol): An open protocol for giving agents tools and data sources beyond their built-in ones. MCP servers you configured for a CLI keep working inside Nix Agents, because the CLI itself is unchanged.
Headless agent: An agent run without a person watching, driven by a script or a queue. The opposite of the interactive session a terminal gives you.
Sandbox: An isolated environment a program cannot escape. Worth being precise: a git worktree separates files and branches, but not ports, databases or a machine — it is not a sandbox.

The words this work is made of.