Subagents

Subagents are autonomous agents created within a larger agent system to handle subgoals or subtasks delegated by a primary agent. They operate semi-independently with their own objectives, perception, and decision-making, but their actions are coordinated by a higher-level controller. The parent may initiate, monitor, or terminate subagents.

In hierarchical or distributed multi-agent systems, subagents enable decomposition of complex tasks, allowing parallel work and scalability. Subagents may be instantiated at runtime based on task demands, and may have restricted scope, access, and authority to prevent unintended actions. Communication usually uses defined interfaces and protocols, exchanging goals, status, and results.

Common techniques include task decomposition, planning with hierarchies, and contract-net-like task allocation. Subagents can be used in robotic systems for modular control, in software agents such as virtual assistants that delegate subgoals, and in simulations and games. They can improve robustness and fault isolation, as failures may be contained within a subagent.

Challenges include coordination complexity, synchronization, risk of misalignment with overall goals, and potential bottlenecks. Design considerations include limiting the subagents' authority, implementing timeouts and watchdogs, ensuring interpretability, and establishing clear termination criteria and recovery procedures.