sub-agents

Specialized AI modules that execute discrete tasks within a broader multi-agent system to boost reliability and speed.

Sub-agents transform monolithic LLM prompts into modular workflows. Frameworks like LangGraph and CrewAI use this architecture to delegate high-level goals to specialized units (think a researcher sub-agent feeding data to a writer sub-agent). By isolating state and memory, these systems reduce hallucination rates by up to 40% in complex reasoning chains. A typical deployment might involve a GPT-4o manager routing 15+ sub-tasks to smaller, faster models like Claude 3 Haiku for parallel execution. This approach ensures that a single failure in a sub-task (like a broken API call) does not crash the entire parent process.

https://www.langchain.com/langgraph