AI

How I use AI has evolved dramatically over the last year. Pi has become my favorite agent harness. It’s small, composable, and easy to shape with extensions. Herdr is the terminal multiplexer around it, keeping agent and terminal processes alive across disconnects. Like all good tooling, both have incredible interfaces to extend the tools.

My setup routes distinct roles to different models and effort levels: explore, executor, verifier, reviewer, oracle, writer, and PR watcher. It also uses reusable skills and centralizes shared configuration across Pi, OMP, OpenCode, Claude Code, Codex, and Amp.

Useful Pi integrations include subagents, LSP, web, MCP and browser access, conversation recall, and a guard for destructive commands. I’ve created some for live voice, better herdr integration, title and progress management and looping.

Herdr adds worktrees, a command palette and workspace picker, pane titles and navigation and layout helpers, plus a workbench where agents can open files in Neovim diffs in LazyGit and run observable jobs. I can run this all from a live Codex voice conversation in the terminal. It’s fun. 🙂

I keep exploring new tools and shape them for engineering tasks as the industry is transformed by AI.