Agentless: Find and Repair Bugs with LLMs

Overview

Agentless is a Python project for using large language models to solve software development issues. Rather than relying on an agent that repeatedly chooses actions, it uses a staged pipeline to locate the likely fault, propose patches, and validate candidates.

The project is particularly oriented toward experiments on SWE-bench. It can suit developers and researchers who want to study or run automated issue repair, but it is not a general-purpose coding assistant setup with no external model requirements.

Key Features

• Hierarchical fault localization, from files to classes or functions and then specific edit locations.
• Generates multiple candidate patches in a diff format for each issue.
• Selects regression tests and can generate reproduction tests for the original error.
• Uses test results to rerank candidate patches before selecting one.
• Provides experiment artifacts for SWE-bench Lite and SWE-bench Verified in the v1.5.0 release.
• Includes repository-structure data and manual SWE-bench Lite classifications.

Use Cases

• Researchers comparing LLM-based software repair approaches on SWE-bench.
• Developers investigating how localization, patch generation, and test-based validation can be separated in an issue-repair pipeline.
• Teams prototyping automated fixes for repository issues when they can provide model API access and run the relevant tests.
• Educators demonstrating an alternative to agent-based workflows for automated programming tasks.

Project Facts

• Language: Python
• License: MIT
• Stars: 2.1k
• Forks: 239
• Topics: agent, artificial-intelligence, llm, software-development
• Archived: false

Getting Started

The repository documents setup with Python 3.11 and a required OpenAI API key:

git clone https://github.com/OpenAutoCoder/Agentless.git
cd Agentless
conda create -n agentless python=3.11
conda activate agentless
pip install -r requirements.txt
export PYTHONPATH=$PYTHONPATH:$(pwd)
export OPENAI_API_KEY={key_here}

See the README and SWE-bench instructions for detailed setup and experiment procedures.

Alternatives

• trae-agent: Trae Agent uses an interactive, tool-using agent for software engineering tasks, while Agentless generates and validates bug-fix patches without an agent workflow.

Considerations

• Running the project requires an OpenAI API key, so usage depends on access to that service and may incur API costs.
• Reproducing the full SWE-bench experiments requires following the repository's separate setup instructions.
• The project is focused on benchmark-style issue repair. The provided description does not establish that it is ready to apply patches safely to arbitrary production repositories.