Subagents are secondary AI agents that a primary agent spawns to handle focused tasks in their own isolated context windows, working in parallel and reporting results back to the parent.
In agentic AI systems, a subagent is an independent agent session created by a parent agent to carry out a specific, bounded task — exploring a codebase, reviewing a change, researching a question — and then return its findings. The defining feature is context isolation: each subagent runs with its own context window, tools, and conversation history, separate from the parent and from any sibling subagents. This keeps the parent's context uncluttered and lets each subagent specialize without interference.
Coordination follows a hub-and-spoke pattern. A parent delegates to subagents and receives their results, but subagents do not communicate with one another — a deliberate constraint that avoids the coordination overhead of every agent talking to every other. When tasks are independent, multiple subagents can run in parallel, often with file-scope or worktree isolation so that overlapping work is detected and serialized. The pattern appears across most major coding agents: 📝Claude Code popularized an explorer, worker, and default subagent model, and 📝OpenAI's Codex offers a close equivalent, configured through TOML files that set each subagent's model, sandbox mode, MCP servers, and skills.
By trading one long-running session for many bounded ones, subagents let a single orchestrator marshal specialized agents in parallel — the dominant pattern for scaling agentic work as tasks outgrow what a single context window can hold.