Comparison
Multi-agent systemvsSubagent
Multi-agent system
you split the work across five agents, spent five times the tokens, and one well-briefed agent did it better.
Several agents with distinct roles cooperating on a task. Genuinely useful for parallel, separable work; also currently overhyped, since a single well-contexted agent frequently beats a committee of them. Adopt when a measured single-agent baseline fails, not before.
Full entry →Subagent
you hand the sub-task to a child with a fresh context and get back only its conclusion, keeping the parent's window small.
A child agent spawned for a bounded subtask with its own fresh context, returning only its conclusion to the parent. The main tool for keeping any single context small, and the standard pattern for parallelising research-style work.
Full entry →