Basically an API wrapping a cyclic graph where rules govern the state transitions / graph traversal (i.e. rules around handing off work between agents and the associated review, rollback and human intervention escalation logic).
It's mostly just to teach myself about multiagent patterns and what blend of "agentic autonomy" and deterministic / human governance gets the best results with the current set of (Anthropic) tools available.
I don't really know what I'm doing w.r.t AI, but having 15 years of industry SWE experience (high-availability distributed systems and full-stack web dev) on top of a fairly-solid CS education I feel like I know what the results of a working system should be and I'm learning a lot about the AI pieces as I go through trial and error.
Generally it feels like there are lots of ways the next generation of AI-assisted coding workflows could work best (beyond just "AI helps write code", I mean) and the results will be as much about the tooling built around the AI bits as it will be the improvements in models / AI bits themselves (barring a theoretical breakthrough in the space).
Trying to figure out what my personal dev workflow will look like in the middle of this evolving landscape is what led to this project, very much a scratch my own itch thing.