Right now the answer is "build it yourself" or "use a free API and hope it doesn't break." If specialist agents (search, code exec, PDF parsing) could be hired on-demand with
per-call billing, you'd see much more composable architectures. The orchestrator focuses on its core logic, and outsources commodity capabilities to paid specialists.
The interesting question is what happens when your stateful agent needs capabilities it doesn't have. Your dev agent needs web search, your research agent needs code execution. Today you build all of that in-house. But imagine a world where your agent just hires a specialist agent for $0.002 per call, gets the result, and moves on.
Microservices solved this for traditional software. The agent equivalent is coming.