Earlier this year I shared a system for coordinating multiple AI agents on research projects. Then I used it on a real one. Here is what changed, what broke, and what I learned when the hardest part turned out to be building a trustworthy dataset.
Plan Orchestrator grew out of a simple frustration: I wanted help from the best AI agents available, but I did not want to babysit them step by step. This is the story of how a demanding private project pushed me to build a more careful, reliable way to use AI for real work.