Agentic Compilation: Reducing LLM Rerun Costs (arxiv.org)

A recent study has introduced a novel approach called "Agentic Compilation" aimed at addressing the significant cost challenges associated with large language model (LLM) reruns during web automation tasks. This phenomenon, deemed the Rerun Crisis, highlights the escalating expenses in token consumption and API latency when LLMs are frequently queried to assess browser states and execute actions. For example, executing a simple 5-step workflow over 500 iterations can cost around 150 USD due to the inherent need for continuous inference. The proposed Compile-and-Execute architecture mitigates this by decoupling LLM reasoning from browser execution, allowing for a one-time LLM invocation that produces a JSON-based workflow blueprint, leading to dramatically reduced costs of under 0.10 USD per workflow. This breakthrough is significant for the AI/ML community as it presents a scalable, cost-effective solution to the automation of repetitive web tasks, breaking the previously linear cost-growth pattern into a more manageable amortized model. By formalizing inference scaling from O(M x N) to O(1), where M is the number of reruns and N is the number of actions, the research ensures that automation remains not just feasible but economically viable. Evaluation in various tasks indicates high success rates of 80-94% for zero-shot compilation, and the modular approach also enhances execution reliability with minimal human intervention, paving the way for more efficient and reliable web automation solutions.