Vendor: princeton-nlp
SWE-agent is a state-of-the-art autonomous software engineering agent from Princeton NLP, pioneering the Agent-Computer Interface (ACI) tailored for LLMs to browse, edit, and resolve real-world GitHub issues.
| Repository | princeton-nlp/SWE-agent |
|---|---|
| GitHub Stars | ★ 20.2k |
| Forks | 2.2k forks |
| Primary Language | Python |
| License | MIT |
| Technical Domain | AGENTS |
$ bash
git clone https://github.com/princeton-nlp/SWE-agent.git
cd SWE-agent
pip install -e .
SWE-agent (princeton-nlp/SWE-agent) is an open-source autonomous software engineering agent developed by the Princeton NLP laboratory. Specifically engineered to excel at the rigorous SWE-bench benchmark, SWE-agent enables large language models to autonomously resolve complex, real-world GitHub issues, bug reports, and pull requests from end to end.
The defining theoretical breakthrough of SWE-agent is the introduction of the 'Agent-Computer Interface' (ACI). Princeton researchers discovered that granting LLMs unfettered access to a standard Linux shell often leads to context explosion and command execution paralysis—models get overwhelmed by massive terminal outputs, struggle to reference specific line numbers, or get stuck in interactive subprocesses. SWE-agent addresses this by designing specialized, LLM-optimized CLI commands (such as customized file viewers with scroll windows, ast-aware search tools, and concise patch replacers) that dramatically streamline model interaction with codebases.
Architecturally, SWE-agent executes entirely within isolated Docker containers to guarantee reproducibility, safety, and deterministic test execution. Featuring a structured reasoning prompt loop and dynamic environment feedback, SWE-agent systematically reproduces bugs, authors targeted patches, executes regression tests, and outputs standard Git patch files ready for human review.
SWE-agent represents the milestone transition of code AI from single-function autocompletion to repository-scale autonomous engineering. When SWE-bench was introduced in 2023, even the most capable frontier models solved less than 2% of real-world GitHub issues when provided with standard command prompts. Princeton researchers diagnosed that the root cause was not an absence of algorithmic reasoning, but an impedance mismatch with traditional human-centric operating system interfaces.
By formulating the Agent-Computer Interface (ACI), Princeton redefined the interaction protocols between LLMs and operating systems. Standard shell commands were replaced with specialized LLM tools: file viewers that display numbered 100-line windows, search commands that summarize hits rather than spewing unmanageable text dumps, and deterministic search-and-replace syntax. This cognitive ergonomics breakthrough multiplied issue resolution rates on SWE-bench by orders of magnitude.
The core insight behind ACI is that tools provided to AI agents must be ergonomically optimized for LLM attention spans and token budgets. By constructing a structured abstraction layer over the operating system, SWE-agent demonstrated that foundation models can diagnose and resolve intricate production defects on par with senior human engineers.
In enterprise open-source and proprietary repository maintenance, SWE-agent monitors incoming bug trackers, reproduces defects, navigates source trees, authors regression fixes, and executes test suites autonomously.
In academic and industrial research, SWE-agent serves as the canonical open-source baseline for benchmarking new frontier foundation models, reasoning strategies, and reinforcement learning code datasets.
In legacy codebase technical debt reduction and library migration, SWE-agent ingests upgrade guides and systematically refactors deprecated API invocations across hundreds of source files.
In automated vulnerability remediation, SWE-agent ingests static analysis vulnerability reports, isolates insecure logic patterns, and authors defensive patches with test harnesses.
Clone the repository and install SWE-agent in editable mode (Docker runtime required):
git clone https://github.com/princeton-nlp/SWE-agent.git
cd SWE-agent
pip install -e .Configure your LLM provider credentials (Claude 3.7 Sonnet or GPT-4o recommended):
export ANTHROPIC_API_KEY="your-api-key"Run SWE-agent to autonomously resolve a real-world GitHub issue and generate a patch:
python run.py \
--model_name claude-3-7-sonnet-20250219 \
--data_path https://github.com/marshmallow-code/marshmallow/issues/1867 \
--config_file config/default.yamlUpon completion, SWE-agent saves full execution trajectories in trajectories/ and outputs a standard patch file ready for review via git apply.
In production practicality and safety, SWE-agent's containerized execution model ensures that all code edits, package installations, and test runs occur strictly inside isolated Docker containers, protecting the host system from unintended side effects. Standard unified .patch outputs allow human maintainers to conduct straightforward code reviews before merging.
Regarding computational cost, solving complex repository-level issues involves iterative reasoning loops, code searching, and test cycles that can consume 500k–1M+ tokens per issue. In production deployments, it is best practice to implement pre-filtering triage heuristics to direct SWE-agent toward well-defined bug reports.
To optimize startup latency, SWE-agent supports pre-building environment Docker images, reducing sandbox bootstrap times from several minutes down to under 10 seconds per task.
With over 20,000 GitHub stars, SWE-agent is recognized globally as a foundational academic and industrial milestone in agentic software engineering. Frontier organizations—including Cognition (creators of Devin), All-Hands AI (OpenHands), and GitHub Next—have incorporated ACI principles into their autonomous architectures.
Princeton NLP continues to push the frontier with SWE-agent Inspector, providing interactive visual step-through debuggers that enable engineers to inspect agent decision chains in real-time.
SWE-agent - Autonomous Software Engineering Agent with Agent-Computer Interface is an open-source AI project developed primarily in Python under the MIT license. SWE-agent is a state-of-the-art autonomous software engineering agent from Princeton NLP, pioneering the Agent-Computer Interface (ACI) tailored for LLMs to browse, edit, and resolve real-world GitHub issues.. SWE-agent represents the milestone transition of code AI from single-function autocompletion to repository-scale autonomous engineering. When SWE-bench was introduced in 2023, even the most capable frontier models solved less than 2% of real-world GitHub issues when provided with standard command prompts. Princeton researchers diagnosed that the root cause was not an absence of algorithmic reasoning, but an impedance mismatch with traditional human-centric operating system interfaces. By formulating the Agent-Computer Interface (ACI), Princeton redefined the interaction protocols between LLMs and operating systems. Standard shell commands were replaced with specialized LLM tools: file viewers that display numbered 100-line windows, search commands that summarize hits rather than spewing unmanageable text dumps, and deterministic search-and-replace syntax. This cognitive ergonomics breakthrough multiplied issue resolution rates on SWE-bench by orders of magnitude. The core insight behind ACI is that tools provided to AI agents must be ergonomically optimized for LLM attention spans and token budgets. By constructing a structured abstraction layer over the operating system, SWE-agent demonstrated that foundation models can diagnose and resolve intricate production defects on par with senior human engineers.
Clone the repository and install SWE-agent in editable mode (Docker runtime required):
git clone https://github.com/princeton-nlp/SWE-agent.git
cd SWE-agent
pip install -e .Configure your LLM provider credentials (Claude 3.7 Sonnet or GPT-4o recommended):
export ANTHROPIC_API_KEY="your-api-key"Run SWE-agent to autonomously resolve a real-world GitHub issue and generate a patch:
python run.py \
--model_name claude-3-7-sonnet-20250219 \
--data_path https://github.com/marshmallow-code/marshmallow/issues/1867 \
--config_file config/default.yamlUpon completion, SWE-agent saves full execution trajectories in trajectories/ and outputs a standard patch file ready for review via git apply.
SWE-agent - Autonomous Software Engineering Agent with Agent-Computer Interface is well-suited for Autonomous GitHub Issue Resolution, Automated Regression Test Repair, Software Engineering Academic Benchmarks, Automated Code Review & Patch Synthesis. With an overall rating of 4.8/5, it offers strong community activity, reliable performance, and easy integration with existing AI pipelines.
In production practicality and safety, SWE-agent's containerized execution model ensures that all code edits, package installations, and test runs occur strictly inside isolated Docker containers, protecting the host system from unintended side effects. Standard unified .patch outputs allow human maintainers to conduct straightforward code reviews before merging. Regarding computational cost, solving complex repository-level issues involves iterative reasoning loops, code searching, and test cycles that can consume 500k–1M+ tokens per issue. In production deployments, it is best practice to implement pre-filtering triage heuristics to direct SWE-agent toward well-defined bug reports. To optimize startup latency, SWE-agent supports pre-building environment Docker images, reducing sandbox bootstrap times from several minutes down to under 10 seconds per task.
Minimal tool for running large language models locally