Orchestra

Anyone who has built with AI agents knows the work doesn't end once you've said what to build. It turns into supervision: “build that,” “review it,” “plan the fixes,” “apply the fixes,” “now review the plan.” Days of that, for a single feature.

I wanted to turn that supervision into reusable instructions: what an agent needs to know, how it checks its work, and when the next phase begins. The challenge was making that process work for any coding task.

1. Working within weekly limits. Orchestra swaps in suitable agents to make better use of multiple subscriptions. The /crosstalk and /throttling scripts let agents spawn one another and check usage by reading their terminals, much as a person would.
2. Managing several projects at once. The /pm agent now operates my dispatchers across projects. It keeps up better than I could: monitors feed it events from every dispatcher and subagent, giving it a live view of the whole fleet.
3. Finding out what each model can do. Company benchmarks can favor their own models, so I run my own A/B tests on real coding tasks. The studies section records how the models compare and which jobs I can trust them with.