Multi-agent review beats single-prompt review

Single-prompt AI code review is generic: naming, semicolons, surface issues. An orchestrator that analyzes the diff and routes to specialist sub-agents running in parallel catches what one generalist misses. Token usage explains most of the performance variance: spending tokens across focused contexts outperforms one wall-of-text prompt.

The orchestrator pattern also solves a cost problem. A Terraform-only PR does not need the frontend specialist; the lead agent skips it, saving tokens. A documentation update needs no specialists at all; the lead handles it directly. This smart routing means you pay for expertise only where it matters. The result is both more thorough (specialists go deep) and more efficient (no wasted tokens on irrelevant perspectives) than a single generalist trying to cover everything.