Get the Nested Subagents Decision Guide

The honest companion to the video on Boris Cherny shipping nested subagents in Claude Code — agents that spawn agents, capped at depth=5, as a way to manage context. The feature is genuinely powerful: each child subagent runs in its own isolated context window and hands back only a clean summary, so your main thread never gets flooded with file dumps, tool outputs, and dead ends. But agents spawning agents form a tree, and trees fan out — one agent calls three, each of those calls three more — so a deep stack can run dozens of agents for a single request, each burning its own tokens in its own window. That is exactly why depth is capped: the limit is a guardrail against a runaway tree quietly torching your usage, not a target to hit. Inside: the honest gain-vs-cost split (isolated context, clean summaries, and parallel work on one side; token multiplication, stacked latency, and harder debugging on the other); the when-to-nest vs keep-it-flat decision (nest when subtasks are independent, context-heavy, and you only need the summary back; stay flat when the task is small, sequential, or you need to steer it); a 3-question test; and the depth ladder decoded (depth 0 flat, depth 1 delegate, depth 2 fan-out, depth 3–4 rare, depth 5 the ceiling) with the one gut check before you go deeper — am I managing context, or just hiding a task I do not understand yet? The takeaway: depth=5 is the ceiling, not the target — most real work lives at depth 1 or 2.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.