Some agents split a big job between subagents. A subagent is another copy of the model doing its own part of the job, a bit like a helper the main agent brings in. When an app says it’s “starting several agents” or “researching in parallel,” each of those agents is a subagent. Some research modes in chat apps work this way, and so do many agents that take on long tasks.
What a Subagent Knows
Each subagent starts with an empty context. It doesn’t see your conversation or the files the main agent has already read, so it knows only what the main agent writes in the instructions it hands over. The subagent reads whatever it needs, does its part, and sends a short summary back:
Only that summary lands in your conversation. Everything the subagent read along the way stays in its own context, so your conversation stays shorter, and a shorter conversation uses less with every message and gets you better answers.
