LangGraph Subagents: Stop AI Agent Context Bloat

If you have built any non-trivial agent, you have hit this wall: the main loop fires three tool calls, each returns a 2 KB blob, the model reads all of it, and by turn five the context window is full of stale search results. The model loses the plot, costs go up, latency climbs.

LangGraph v0.4 (April 2026) shipped native subagents — first-class child graphs whose internal tool traffic never bubbles up to the parent. The parent agent sees only the final summary the subagent returns. Context stays small, the agent stays focused, and you can specialise each subagent with its own prompt and toolset.