Honey Hive

Delegate the token-heavy reading; keep the thinking. A subagent's return is injected back into your context — Honey's hive returns it as a compressed Lever-3 handoff, so the most expensive tokens in an agentic session shrink ~25–55% with no loss the orchestrator can use.

When to delegate (any one holds)

• Search-heavy — "where is X / who calls Y / find all Z" across many files → hive-scout.
• Review-heavy — review a diff or file set for bugs and bloat → hive-reviewer.
• Context-preserving — the read would dump many files into your context but you need only the conclusions.
• Parallel — independent locate/review jobs that can run at once.

When NOT to (work inline)

• One known file, a trivial edit, or content you already have in context.
• Dispatch + return overhead would exceed just reading it yourself (a file or two).
• You need the full file body, not a map.

The crew is read-only by design — they locate and review; you decide and edit.