The Pantheon: building my own AI agent harness in public

I've been running Claude everywhere, doing a thousand things at once. The more I did it, the clearer it got: watching an agent code next to you doesn't scale. The harness has to become architecture. This is the Pantheon — a set of plugins, each named for a god, each able to run on its own or together: Hive — spec-driven development Consus — edit the docs and diagrams the agents produce, and make the decisions Mnemosyne — memory you can actually A/B test Portunus — a secret broker, so keys stay out of the model Minerva — planning plus an orchestrator, ticketing (wrapping Multica), T3 Code for the code view, and a UI layer (Janus) I'm just starting. I've built three or four orchestrators before this one. Every one was hard-coded to me and coupled together. This time I pulled it apart so each piece stands on its own. *The goal isn't to watch agents work. It's to make the decisions they can't.* Full overview (10 min). Each plugin gets its own video as it ships — Portunus is first. Portunus is open source: https://github.com/mdostal/portunus Hive is open source: https://github.com/firefly-events/plu... #AI #AIAgents #ClaudeCode #BuildInPublic Everything I'm releasing lives at tools.mdostal.com. Posted with flayr.social — try it: flayr.social/r/4yqhaf/yt