LangChain introduces dynamic subagents in Deep Agents, a feature that lets AI agents write short orchestration scripts (JavaScript) to dispatch and manage subagents programmatically rather than through sequential tool calls. This approach solves two key problems: deterministic coverage at scale (e.g., processing 300 pages via a loop instead of 300 tool calls) and reliable complex orchestration with fan-out, branching, and multi-phase pipelines. The post covers six orchestration patterns — classify and act, fanout and synthesize, adversarial verification, generate and filter, tournament, and loop until done — with live traces for each. Setup requires installing the QuickJS middleware and using the CodeInterpreterMiddleware with create_deep_agent, or using the dcode terminal agent which ships with the feature enabled out of the box.
