Agent Analysis

A comprehensive toolkit for analyzing Software Engineering (SWE) agents, focusing on performance evaluation, feature analysis, and benchmarking.

Overview

This project provides tools and utilities for:

• Analyzing agent performance on software engineering tasks
• Computing various metrics (code, dependency, error, instance, patch, type)
• Evaluating performance gaps between different agent implementations
• Processing and analyzing data from OpenHands and SWE-bench

Project Structure

• analysis/: Core analysis modules

  • features/metrics/: Various metric implementations for agent analysis
  • models/: Data models for OpenHands and SWE-bench
  • performance_gap.py: Performance gap analysis utilities
  • usage.py: Usage analysis tools

• notebooks/: Jupyter notebooks for analysis and visualization

  • condenser_results.ipynb: Analysis of condenser results
  • localization_metrics.ipynb: Metrics for code localization
  • performance_gap.ipynb: Performance gap analysis

Requirements

• Python ≥ 3.12
• Dependencies are managed through Poetry

Installation

1. Ensure you have Poetry installed
2. Clone this repository
3. Run poetry install to install dependencies

Usage

The toolkit can be used either through its Python modules or via the provided Jupyter notebooks for interactive analysis.

License

This project is licensed under the MIT License - see the LICENSE file for details.