In plain words
Subagents are agents that work on delegated parts of a broader task. A coordinating agent or workflow gives each subagent a bounded assignment, then uses the returned results to continue or complete the overall job.
A closer look
Subagents are useful when work can be divided by specialty or pursued in parallel. One might research sources while another reviews code or checks calculations. The coordinator decides what context each subagent receives, whether their work can happen simultaneously, and how their outputs are combined.
Good delegation requires clear scope, enough context, and an explicit expected result. More subagents do not automatically improve an outcome: they add communication cost, may duplicate work, and can produce conflicting conclusions. The coordinating system still needs to verify results, resolve disagreements, and maintain responsibility for the final output.
In practice
A lead research agent asks one subagent to find primary sources, another to compare competing claims, and a third to check the draft’s citations. The lead combines their findings into one answer.
A useful distinction
A subagent is not necessarily a smaller, weaker, or separately trained model. It may use the same model as the coordinating agent with different instructions, context, tools, or permissions. A handoff can transfer control to another agent, while a subagent often returns its result to the coordinator.