A subagent in 📝Hermes Agent is an isolated agent spawn with its own conversation, terminal, and tools, dispatched for delegation or parallel work.
A subagent is launched by a parent agent — sometimes called the orchestrator — to handle a discrete task in isolation from the parent's context. It runs in its own sandbox backend with its own tool set and conversation history, and returns a structured result when the work is done. The parent does not pay context cost for the subagent's intermediate reasoning, only for its final return value, which keeps the parent's window focused on coordination rather than execution detail.
Subagents can themselves spawn subagents when granted the orchestrator role, with configurable max_spawn_depth to bound nesting. The pattern composes: a top-level agent splits a large task into parallel subtasks, each subagent further splits its piece, and the orchestrator collates results back up. Hermes' multi-agent constructs are built on this primitive — Kanban workers, /goal delegation, and Autonomous Curator review passes are all subagent dispatches in disguise.
Isolation is the load-bearing property. The cost trade-off is straightforward: an inference call per spawn plus the subagent's own token usage, in exchange for parallelism and a tighter parent context. Most multi-step pipelines in Hermes prefer the subagent pattern over inline tool calls for tasks that would otherwise saturate the parent agent's context budget — a deliberate handoff of complexity from one window to many.