Minari

This is a personal prototype for multi-agent workflows on a DAG. You put agents in a graph, give each a role and a place in a hierarchy, then run the team with a person still in the loop.

It isn't a shipped product. What exists so far is the editor, the hierarchy, and pause and playback.

I needed a run I could stop and look at. Linear prompt chains hide branching, and a fire-and-forget request hides the person who might need to step in.

I built the editor on a DAG, with each agent given a role and a place in a hierarchy so you can staff the team and debug it. Pause and playback let someone take the run back. That's as far as the prototype goes. I'm not claiming a marketplace or a hardened deploy path.

I used a DAG instead of a transcript because branching was the thing I was trying to build, not an edge case.

Roles and hierarchy make it possible to staff the team and debug it. A flat swarm of prompts is harder to read. Human-in-the-loop review, plus pause and playback, lets a person take the run back. I'm not claiming a marketplace or a hardened deploy path. This is still a prototype.

It's still a prototype. What I'm trying to prove is fairly narrow: you should be able to see the team, assign the seats, and pause the run.

If a person can't get back into the graph, the graph isn't that useful.