Subagents are isolated instances of an AI agent, each with its own context window, tool access, and system prompt, that a parent agent delegates a specific task to and receives back only a finished result from, keeping the delegated work's intermediate detail out of the main conversation.
What does isolation actually buy you?
Every exploratory step inside a subagent’s run, the searches that missed, the file reads that turned out irrelevant, stays inside that subagent’s own context and never reaches the parent. The parent conversation only ever sees the finished result, which keeps a long running task from filling up with noise from every subtask it delegated.
Where do subagents show up in production systems?
They are the delegation layer of deep agents: a planner decides a subtask is separable, spins off a subagent scoped to exactly that piece, and moves on once the result comes back. In a broader multi-agent system, subagents are typically the workers in an orchestrator-worker pattern, each handling one kind of subtask without needing to know what the others are doing.
When is delegating to a subagent actually worth it?
When a subtask is genuinely separable, has a clear boundary, and either needs focused context or can run in parallel with other subtasks. A short task with one obvious next step does not need the overhead; that is what a single AI agent running its own loop is for.
Frequently asked questions
Why not just let one agent do everything in its own context?
A single context fills up fast: every search result, every file read, every intermediate attempt stays visible and competes for the model's attention. A subagent does that exploratory work in its own isolated context and returns only the finished answer, so none of the noise reaches the parent.
Can subagents run at the same time?
Yes, when the subtasks do not depend on each other's output. Running independent subagents in parallel is one of the main practical reasons to use them instead of one agent working through the same subtasks in sequence.
Do subagents need their own separate tools?
Not necessarily the same tools as the parent. A subagent is usually scoped to only the tools its specific subtask needs, which narrows its failure surface and makes its behavior easier to predict than giving every agent access to everything.