Subagents are specialized AI agents purpose-built for single, well-defined tasks, used to combat context pollution in monolithic AI agents. Two patterns exist: explicit user-defined subagents (static, predictable, reusable, as seen in Claude Code) and implicit on-the-fly subagents (dynamic, flexible, stateful, as seen in Poke.com). Explicit subagents offer control and predictability but are rigid; implicit ones are flexible but harder to debug. Both approaches improve LLM performance by providing focused context, specific instructions, and limited toolsets. Predefined subagents suit enterprise workflows while dynamic ones fit B2C applications. Reliability remains a challenge, but decomposing tasks into subagents makes them easier to test in isolation.
