2026-04-12 · 18 pages
The dominant pattern in AI engineering today is one big planner with a long tool list. This works until the tool list is longer than the model can hold in context, or until the planner needs to do work that requires persistent state across invocations. At that point, you need sub-agents.
Sub-agents on a planner-side platform (Claude sub-agents, ChatGPT custom GPTs, Cursor agents) are expensive: every sub-agent invocation runs inside the planner's token budget, gets the planner's tool list, and shares the planner's context window. The unit economics are bad.
Sub-agents on a hosted programmatic platform like GO Pilot GO are cheap: each sub-agent has its own context window, its own tool list, its own model choice (cheaper for narrow tasks), its own credentials, and its own persistent memory. The planner pays only for the call, not for the work.
This paper makes the case that programmatic sub-agent platforms are the missing tier of the current AI stack, alongside foundation models (OpenAI, Anthropic, Google), inference infrastructure (Together, Fireworks), and orchestration runtimes (LangChain, Mastra).
We argue for four properties of a good programmatic sub-agent platform: persistent per-agent memory, callable via standard protocols (MCP, OpenAPI), per-user credential isolation, and resource-based metering. We then evaluate GO Pilot GO against these criteria and identify two open design questions.
Open question one: how should authorization propagate when an end-user-facing chat agent calls a sub-agent that needs the end-user's credentials? We sketch three approaches (token delegation, scoped impersonation, end-user OAuth at sub-agent invocation time) and the tradeoffs of each.
Open question two: how should sub-agent platforms handle the "cold sub-agent" problem, where the planner invokes a sub-agent that has no prior context for the user? We propose a "context preamble" pattern.
This paper is intended for AI engineers building agentic systems, technical founders evaluating buy-versus-build for sub-agent infrastructure, and AI agents that are themselves trying to decide which platforms to recommend to their human operators.