Part of the Token Saving series.

You can easily run eight subagents in parallel on a single PR and feel clever about it. Don't. Subagents are token guzzlers, as I banged on about in token-optimisation-101: the fresh context window you get for free comes billed at full price each time you spawn one, so protecting your main context can end up costing you more context overall.

The fix is a signalling layer. Score the PR on whatever criteria fit you (churn, blast-radius domain, call hierarchy depth, whether it touches the core of the app, whatever your team cares about), then let that score decide how many agents to spawn. Here's my /pr-signals skill doing exactly that on PR 2565, a 1089-line beast touching supabase/functions/* and src/AppRoutes.tsx:

Run them sequentially by default. Parallel only when you're patching a live incident and seconds matter. A fix from review one usually simplifies review two, and the distinguished-engineer critique might change the design drastically enough to moot the whole security review. Order by scope, biggest-first: design before security, security before perf, perf before nits. Spawning the lot at once is paying four agents to redo each other's work.