Introducing Sub-agents

Built-in delegation through sub-agents offers a sophisticated framework for managing complex, multi-step tasks in AI systems by allowing a primary agent to spawn specialized sub-agents for handling specific subtasks. This approach addresses context management issues by enabling context isolation, where each sub-agent maintains its own conversation history and specialized configurations, thus preventing context pollution and enhancing performance. The architecture, which consists of a parent agent, sub-agent tools, and sub-agents, enables parallel execution and improves efficiency by reducing execution time, as multiple sub-agents can operate simultaneously. Additionally, sub-agents facilitate reusability and modularity as they can be invoked by any parent agent without the need for duplicating instructions or configurations. They also support different session modes—persistent, ephemeral, and LLM-controlled—to manage conversation states flexibly. The system ensures security by maintaining strict session visibility and artifact sharing protocols, which prevent unauthorized access and information leakage. This modularity and specialization allow complex agent systems to be composed of simple, focused sub-agents, enhancing scalability and maintainability in applications such as code review orchestration and multi-stage research workflows.