Agentic Compilation: Mitigating the LLM Rerun Crisis for Minimized-Inference-Cost Web Automation

Researchers propose a Compile-and-Execute architecture that reduces LLM-driven web automation costs from $150 to under $0.10 per workflow by decoupling reasoning from execution. Instead of continuous inference loops, a single LLM call generates a deterministic JSON blueprint that a lightweight runtime executes without additional model queries, achieving 80-94% zero-shot success rates.

The paper addresses a critical economic barrier in deploying LLM-based web agents at scale. Continuous inference architectures—where models repeatedly evaluate browser state and select actions—create a compounding cost problem: token expenditure grows linearly with task iterations and sequential steps. For practical enterprise workflows, this makes automation economically unviable despite theoretical capabilities.

The Compile-and-Execute paradigm shifts the computational model fundamentally. Rather than treating the LLM as a runtime decision-maker, it relegates the model to a one-time planning phase. A DOM Sanitization Module creates a token-efficient semantic representation of web pages, which the LLM processes once to emit a deterministic workflow specification. A deterministic runtime then executes this blueprint without further model inference, eliminating the variable cost component entirely.

This approach carries substantial implications for the broader AI automation ecosystem. The shift from O(M × N) to O(1) amortized inference scaling directly enables business models previously impossible—large-scale data extraction, form automation, and web intelligence at competitive costs. The 80-94% zero-shot success rates with minimal human-in-the-loop patching suggest practical deployability without extensive fine-tuning or prompt engineering overhead.

The findings underscore a critical principle emerging across AI infrastructure: not all reasoning tasks require continuous model access. By decomposing workflows into planning and execution phases, developers can dramatically reduce inference costs while maintaining reliability. This architectural insight extends beyond web automation, suggesting broader optimization opportunities in agentic systems that could reshape cost economics across enterprise AI deployments.

• →LLM web agents reduce per-workflow costs from ~$150 to under $0.10 using compile-once-execute-many architecture instead of continuous inference loops
• →Single-shot LLM compilation generates deterministic JSON blueprints that lightweight runtimes execute without additional model queries, achieving O(1) amortized scaling
• →Zero-shot success rates of 80-94% across data extraction and form-filling tasks demonstrate practical viability with minimal human-in-the-loop intervention
• →Cost reduction enables economically viable automation at scales previously infeasible, particularly for repetitive enterprise workflows at 500+ iterations
• →DOM Sanitization Module creates token-efficient semantic representations, enabling deterministic execution without real-time model decision-making