I run Suhail, my own Claude Code orchestrator, against real repos most days. The first time I timed a full run (an indexer plus a trivial two-part plan), it took 22 minutes. My instinct was to blame the model: swap to a faster tier, shave a few seconds per call, ship it.
That instinct was wrong. I pulled the actual dispatch counts, and the bottleneck wasn't model latency at all. It was how many times I was cold-starting a subagent to re-read context it had already seen.
The real cost: dispatch count, not tokens-per-second
Each "Part" in that run fired five role dispatches: researcher, planner, reviewer, auditor, and the coder. Four of those five paid a full cold start, fresh context, re-reading the same research bundle, the same plan, the same diff, before doing maybe 30 seconds of actual reasoning. The model was fast. The pipeline was slow because I'd designed it to re-derive the same context five times per unit of work.
Once I started treating dispatch count as the metric to optimize instead of "which model is faster," the fixes were mechanical:
