AgentCity

AgentCity is an LLM Agent-driven automation platform for urban computing research, built on top of the LibCity framework. It leverages AI agents to automate the full lifecycle of traffic prediction research — literature search, model migration, hyperparameter tuning, and benchmarking — automatically porting state-of-the-art models from published papers into a unified framework for reproduction and evaluation.

Key Features

• Literature Search: Retrieve and organize the latest traffic prediction papers by keywords, year range, and target conferences.
• Model Migration: Migrate open-source model implementations from papers into the LibCity unified framework, generating compatible model code.
• Hyperparameter Tuning: Perform automated hyperparameter search and optimization on migrated models.
• Benchmarking: Run evaluations on standard datasets and produce comparison reports with metrics.

Project Structure

AgentCity/
├── agents/                  # AI agent core modules
│   ├── core/                #   Orchestrator, pipeline, config
│   ├── literature/          #   Literature search workflow
│   ├── migration/           #   Model migration workflow
│   └── tuning/              #   Hyperparameter tuning workflow
├── Bigscity-LibCity/        # LibCity framework (models, datasets, configs)
├── frontend/                # Web frontend (HTML/JS/CSS)
├── data/                    # Data storage
├── datasets/                # Datasets
├── documentation/           # Migration docs and verification reports
├── repos/                   # External paper source code repositories
├── server.py                # FastAPI backend server
├── batch_runner.py          # Batch runner (migration + testing)
├── batch_migration.py       # Batch migration script
├── claude_client.py         # Claude API client
└── requirements.txt         # Python dependencies

Installation

Steps

1. Clone the repository:

git clone <repository-url>
cd AgentCity

2. Install Python dependencies:

pip install -r requirements.txt

3. Set environment variables:

export ANTHROPIC_BASE_URL="<your-api-base-url>"
export ANTHROPIC_API_KEY="<your-api-key>"

Alternatively, you can place these variables in a .env file in the project root.

If you want to use other models, install cc-switch to switch between different model providers.

4. Prepare datasets (optional, required for model training and testing):

Download LibCity datasets and place them under Bigscity-LibCity/raw_data/.

Quick Start

Start the web server

uvicorn server:app --host 0.0.0.0 --port 8000

Then visit http://localhost:8000 in your browser to access the dashboard.

Run batch migration and testing

python batch_runner.py --catalog test_flow.json

Usage Guide

Frontend Usage

After starting the web server, open http://localhost:8000 in your browser to access the LibCity Agent Dashboard. The frontend provides the following modules:

1. Literature Search

• Enter keywords (e.g., traffic prediction, spatio-temporal).
• Select a year range (this year / last year / all / custom range).
• Filter by target conferences (ICLR, ICML, NeurIPS, KDD, ICDE, etc.).
• Click search to retrieve matching papers. Results include direct PDF access links.

2. Model Migration

• Select target papers from the literature search results.
• Trigger a migration task. The agent will automatically:
  • Clone the paper's source code repository
  • Analyze model architecture and data flow
  • Generate LibCity-compatible model code
  • Register model configurations
• Migration progress and stage logs are viewable in real time on the dashboard.

3. Hyperparameter Tuning

• Select a migrated model and launch a tuning task.
• The agent performs automated hyperparameter search and outputs the optimal configuration.

4. Job Management

• The dashboard includes a job queue panel showing all running and completed tasks.
• Each job displays its label, status, and associated metadata.

Command-Line Usage

Batch Migration and Testing

Use batch_runner.py to process papers from a catalog file:

python batch_runner.py --catalog test_flow.json --output benchmark_results.csv

Batch Migration

Use batch_migration.py to run model migration only:

python batch_migration.py --catalog migration_flow.json

Verify Migration Results

Ensure LibCity datasets have been downloaded to Bigscity-LibCity/raw_data/, then run:

cd Bigscity-LibCity
python run_model.py --task traffic_state_pred --model <model_name> --dataset <dataset_name>

Publication

This work accompanies the following paper received in the Frontiers of Computer Science special column “Code & Data in Earth Science”:

LibCity2025: Recent Advances and Unified Evaluation for Urban Traffic Prediction, Frontiers of Computer Science, 2026.