A subagent is an AI agent launched by another agent to do one bounded job — search these forty files, test this branch, summarise these papers — and hand back a result. The parent keeps the plan. The child does the digging and returns an answer rather than the pile of material it waded through.
The reason this pattern exists is the context window. An agent working a long task fills its context with intermediate junk: file listings, failed attempts, tool output nobody needs again. Delegating that work to a subagent means only the conclusion comes back, so the parent stays coherent over far more steps than it otherwise could. Several subagents can also run at once, which is why agentic tools have become dramatically faster at tasks that parallelise.
Why it matters at your desk. It explains both the capability jump and the bill. When a coding tool suddenly handles a repository-wide refactor it used to fumble, subagents are often why. When your usage allowance disappears faster than expected, subagents are also often why — each one spends tokens, and a single instruction can fan out into a dozen parallel runs. The same prompt can cost wildly different amounts depending on how much the agent decides to delegate.
What to watch for: subagents make an agent's work harder to follow. The parent reports a tidy summary, and the reasoning that produced it happened somewhere you did not see. For anything consequential, ask the tool to show the intermediate steps rather than trusting the digest.