Orchestration

Coding agents are brilliant and unreliable in roughly equal measure. They start strong, wander off, forget what they were doing, and declare victory on work they never finished. Orchestration is my answer: a Claude Code plugin that lets me hand an agent real work and get it back actually done — properly, not confidently half-done.

Interruptions stop mattering. Every task gets scoped, planned, executed, and verified — in that order, every time — against a plan that lives on disk instead of in the agent's memory. So when that memory resets mid-build, the agent reads the plan and picks up exactly where it stopped. No re-explaining, no starting over, no quiet drift from what was agreed.

And nobody grades their own homework. Before anything counts as done, a rival AI from a different model family reviews the work and has to sign off. I stay in charge the whole way: plans land in my editor for review, and nothing ships until I approve it — approve or send it back, the agent doesn't get a vote. When a job should run elsewhere, the work moves to a second machine without two agents ever stepping on each other.

None of it works by force — no hard gates, no blocked tools. Discipline as prompts, not police. And the proof is the loud part: the site you are reading was built this way. This page began as a step in a plan and did not ship until a rival model signed off.