Subagents: the unit of parallelism that actually pays off
Subagents earn their keep for parallel search, large-context lookups, and anything that would pollute your main context. Delegate lookups, never delegate understanding.
The first month I used Claude Code, I tried to delegate everything to subagents. The second month, I stopped delegating anything. The eight-month version is a clearer heuristic: delegate to protect context, never to skip thinking.
Where subagents win: searching across a large codebase, running parallel exploration of unrelated questions, and any task that returns a small summary from a large input. An Explore agent that reads 40 files and tells me where a hook is defined saves me 40 file reads in my main thread. A Plan agent that drafts an implementation strategy without polluting my working context is genuinely useful.
Where they lose: anything that requires you to understand the answer to make the next decision. Delegating 'fix this bug' is a trap, because the subagent's summary describes what it did, not whether the fix was right. The trust gap is the cost. If the next step is high-stakes, drive it yourself and verify the diff.
The parallelism unlock most teams miss: launch independent subagents in parallel from the main thread. Spinning up an Explore agent for code search at the same time as a Plan agent for architecture sketching is two-minute work that would otherwise be ten-minute sequential context-switching. The harness handles it; you just have to remember to do it.