Prefer feature-specific subagents over generic 'qa'/'backend' agents, and spawn before context pollution
MCP Directory·medium signal
Boris Cherny's 2026 rule of thumb: feature-specific subagents beat general role agents ('qa', 'backend engineer') because specificity buys better tool selection and tighter context. The higher-leverage habit is to spawn a research subagent the moment a task would pollute your main context — do the 20 file reads and 12 greps inside the subagent so your main session sees only the final report, then plan with a clean context. This is the single biggest move for keeping main contexts small.