I like this mental model best: brain = LLM body = agent harness brain + body = agent key thing to understand is that tools like Claude Code and Codex are not JUST the agent harness they CONTAIN the agent harness the harness is the software that implements and manages the agent loop. here's a concrete example: Codex (the app/CLI) is built on the Codex Core agent harness, which is open source. The agent runtime is implemented in Rust in openai/codex/codex-rs/core (on github) Claude Code's underlying agent harness is not open source. However, the Claude Agent SDK is open source and provides programmatic access to the same underlying runtime used by Claude Code The Codex or Claude Code App (or CLI) actually bundle both the AGENT HARNESS and the INTERFACE into the agent The "agent" here is the combination of HARNESS + LLM This is all jargon, at the end of the day. But as an educator its important for me to know what I'm talking about when I say "agent" or "agent harness" This was inspired by Theo's recent content explaining harnesses, which GREATLY helped develop my thinking around this topic: https://x.com/theo/status/2117213872348348429

If you follow Theo on YouTube or X then you know about t3 code (If not then I recommend checking it out. maybe you'll like it. AND interestingly, the code is open source: https://github.com/pingdotgg/t3code) The thing is, t3 code isn't for me. As you know, I use and love CLI agents + tmux (claude code, codex cli, etc..). And t3 code is GUI only. BUT I've recently been feeling the pain of managing many agents SO I created my own version of t3 code. I called it tmux-agent-plane. It's a CLI + agent skill The way it works is with a CLI (called `tmux-agent-plane`) and an agent skill The CLI sits on top of tmux, i.e. its basically an opinionated wrapper around the `tmux` CLI. The agent skill teaches my agents how to use the CLI. There's some additional overhead, since the agents are often now needing to load this skill and use the CLI in order to do their everyday work. But for me, its completely worth it. This has really been life changing! Here's where you can get it and try it out: https://github.com/zazencodes/tmux-agent-plane There's also another video demo in the README that shows how it works Fair warning: I'm shipping a lot of changes for this, and I haven't got it working yet with Pi or OpenCode. Of course, I'd encourage you to play around with it. And I'm happy to (have my agents) review and merge in your PRs on the project!