The Rise of Subagents (www.philschmid.de)

AI developers are increasingly adopting "subagents" — purpose-built, single-task agents orchestrated by a higher-level controller — to avoid the context pollution and brittleness of monolithic agents. Subagents bundle a name, capability description, system instructions, a limited toolset, and an isolated context window; an orchestrator delegates work to them so each subagent can focus on a clear, repeatable goal. Examples include Claude Code (explicit, user-defined subagents) and Poke.com (implicit, on‑the‑fly subagents). There are two main patterns: explicit subagents are statically defined (in code or files), list their tools up front, and are predictable, reusable, and easy to test, but rigid to update and harder to scale to hundreds of specialties. Implicit subagents are dynamically instantiated by an orchestrator (e.g., via a send_message_to_agent call), select tools from a pool based on natural language, and can be stateful across runs — offering flexibility and smoother multi‑step workflows but making behavior less predictable, harder to debug, and prone to misassignment. Practically, predefined subagents fit structured enterprise workflows while dynamic agents align with consumer-facing use cases. Across both, “context engineering” — giving the model the right info and tools at the right time — is key: subagents improve performance, testability, and cost-efficiency, but reliability and orchestration remain the core engineering challenges.