LLM-as-AgentCompiler: Towards an extremely easy paradigm for industry-level agent deployment

ReAct agents are increasingly deployed to translate LLM capabilities into real applications, but their deployment is limited by high cost and instability, as they are built for multi-turn user interaction rather than reliable, low-cost automation. Recent work compiles ReAct agents into cheaper, more stable workflows, but existing methods are often complex or domain-specific. We propose LLM-as-AgentCompiler, a simple, general three-role agentic framework that uses coding agents to compile ReAct agents into cheaper, more stable workflows for high-throughput deployment. Evaluated on two benchmarks (six task domains) with no human supervision, it achieves roughly $2\times$–$33\times$ token reductions with no significant accuracy loss on five of six domains, plus one identified counterexample (retail under a weak residual model, which we analyze). We show that even a simple, pure-LLM framework with no expert priors can automatically (1) identify the deterministic decisions and computations in the memories, tools, and traces of ReAct agents across diverse domains, and (2) compile them into more static workflows with sparse, relaxed LLM nodes. We view LLM-as-AgentCompiler as an initial proof of concept toward this paradigm; establishing its generality on more open-ended domains remains future work.