Compiling Agentic Workflows into LLM Weights Cuts Costs by 100x

A new arXiv paper (2605.22502) proposes compiling agentic workflows directly into LLM weights via fine-tuning, creating 'subterranean agents' that achieve near-frontier quality at two orders of magnitude less cost. Current agent orchestration frameworks (LangGraph, CrewAI, Google ADK, OpenAI Agents SDK, Semantic Kernel, Strands, LlamaIndex) exceed 290,000 GitHub stars but rely on external orchestrators that inject instructions and routing decisions each turn. Prior work (SimpleTOD, FireAct, SynTOD, WorkflowLLM, Agent Lumos) demonstrated the technique, yet developer adoption has favored orchestration. The approach addresses context window consumption, frontier model dependency, and exposure of proprietary procedures to third-party providers.