Agents.md speaks Unix, and you should too

A practical manifesto for treating coding agents as Unix-native collaborators, arguing that a command-line-centered, file-and-process-based environment gives agents predictable, composable interfaces to read, modify, test, and deploy configuration. Years of curated dotfiles evolved into a reproducible chezmoi/mise layout where agents can propose and apply improvements; Unix primitives (files, streams, exit codes) and tool composition (git, ripgrep, small programs) let agents work without bespoke integrations. Clear, machine-checkable commands, useful error messages, and non-interactive behavior are essential so agents do not get stuck; human review remains the bottleneck and must be supported by diffs, validation checks, and explicit instructions about which copy of a file to change.

AGENTS.md is presented as an open Markdown convention for project-specific instructions that tell agents where durable edits belong, how to validate them, and what to preserve - examples include editing chezmoi sources rather than ~/.config, running predefined checks, and previewing diffs. Complementary Skill directories hold procedures; a custom Python CLI (dot) provides typed, inspectable commands (dot doctor, dot agent context, dot trust, dot secret) and enforces checks. Practical toolchain examples (Zizmor, Lefthook, Cosign, LazyGit, Zellij, Neovim, Claude Code) and a workspace rule - one session per project, one tab per task, one pane per program - emphasize coordination, reviewability, and explicit trust/permission choices.