noun · AI concept
Subagent
also called helper agent
A helper with its own empty backpack. The main AI sends it off with one job and gets back a short report.
An agent launched by another agent to handle a focused task in its own separate context, returning only a result.
The main agent starts another agent with a narrow task and a fresh context. The helper works through its own steps, then returns a result, usually a summary, so the long trail of its work never fills the parent's context window.
Real-life example
You ask an assistant to find every place in a large codebase that handles payments. It sends a helper to search. The helper opens dozens of files and rereads them over many steps. The main assistant only receives a list of five relevant spots, so your main conversation stays clear.
Where the analogy breaks
-
A subagent is not a shared brain. It sees only what it is handed, so a vague instruction gets a vague result.
-
What comes back is a summary, and summaries lose detail. Something the helper noticed but did not report is gone for the main agent.
-
More agents means more total token use. The saving is in the parent's context, not in the overall bill.