ZhuLong LLM Agent Automates EDA Scripting with Offline API Exploration

Electronic Design Automation (EDA) often requires scripting with highly specialized and frequently undocumented APIs, creating a significant bottleneck for engineers. A new LLM-based agent called ZhuLong aims to streamline this process by automating EDA scripting, even for complex and proprietary tools. ZhuLong integrates several key capabilities: it can retrieve relevant API information, inspect available documentation, and execute code within a sandbox environment. Crucially, it features an offline API self-exploration mechanism that allows it to infer the behavior of undocumented APIs through systematic counterfactual experimentation, effectively learning how to interact with the tools without explicit instructions. Evaluations on real-world EDA tasks showed ZhuLong achieving a 78.5% Pass@1 rate, a substantial improvement over pure LLM baselines. The sandbox execution was identified as the primary driver of performance, with the self-exploration mechanism further boosting accuracy and reducing the number of tool calls needed per task. This advancement promises to significantly reduce the manual effort and expertise required for EDA scripting.