AnalogAgent: Self-Improving Analog Circuit Design Automation with LLM Agents

This repository contains the implementation of AnalogAgent, an agentic framework for automated analog circuit design.

🎉 Accepted at KDD 2026 (AI4Science Track), selected for Oral presentation.

📄 Paper: AnalogAgent: Self-Improving Analog Circuit Design Automation with LLM Agents

Overview

AnalogAgent integrates:

• Multi-Agent Systems for execution-driven refinement
• Self-Evolving Memory for cross-task knowledge reuse
• Adaptive Design Playbook for structured generation guidance

🚀 Reproducibility & Artifacts

To support the verification of our methodology and results, we have released the AnalogAgent-Artifact repository. This includes the multi-agent system(MAS), the Self-Evolving Memory (SEM), and the complete evaluation suite.

📦 Repository Structure

The current repository contains the following key components for the refinement loop:

• main_run.py: The central entry script that orchestrates the end-to-end refinement loop (Generate → Execute/Check → Optimize → Curate Rules).
• agents.py: Defines multi-agent roles, including the Code Generator and Design Optimizer.
• curator.py + playbook.py + playbook.json: Implements the Self-Evolving Memory (SEM). Handles rule filtering, conflict checking, deduplication, and persistent storage of design heuristics.
• problem_set.tsv: Benchmark task metadata containing standardized analog circuit design specifications.

🛠️ Execution Environment

All experiments were conducted and verified under the following settings to ensure consistency and performance:

• Operating System: Ubuntu 20.04.6 LTS
• Python Version: 3.10 (A dedicated Conda environment is highly recommended)
• Circuit Simulators:
  • ngspice: Must be installed and accessible in your $PATH.
  • PySpice: Required for Python-SPICE interfacing.
• ML Runtime Stack:
  • Local Inference: Supports vLLM-based serving (optimized for NVIDIA RTX A6000 / CUDA 12.4).
  • Cloud Inference: Supports Gemini, GPT, and Qwen API integrations.
• Hardware: Experiments were executed on a dual-GPU workstation with NVIDIA RTX A6000 (Driver 550.54).

💻 Quick Start: How to Reproduce

You can verify the AnalogAgent methodology by running the following command to initiate an automated design task:

# 1. Clone the repository
git clone https://github.com/TheWind-upBird/Analogagent.git
cd Analogagent

# 2. Install dependencies (Requires Python 3.10)
pip install openai google-generativeai PySpice optuna pandas python-dotenv matplotlib numpy
# Also requires ngspice installed and available in your $PATH (see Execution Environment).

# 3. Run the end-to-end refinement loop (Example: Task ID 1)
python main_run.py \
  --model gpt-5 \
  --api_key YOUR_API_KEY \
  --task_id 1 \
  --num_per_task 1 \
  --num_of_retry 1

Status

Note: This is a minimum viable product (MVP) . A more comprehensive and detailed version of the code will be uploaded subsequently.