SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

By John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Adriano Lieret, Shunyu Yao, Karthik Narasimhan, Ofir Press. Neural Information Processing Systems, 2024

1,448 citations recorded by the source corpus.

Language model (LM) agents are increasingly being used to automate complicated tasks in digital environments. Just as humans benefit from powerful software applications, such as integrated development environments, for complex tasks like software engineering, we posit that LM agents represent a new category of end users with their own needs and abilities, and would benefit from specially-built interfaces to the software they use. We investigate how interface design affects the performance of language model agents. As a result of this exploration, we introduce SWE-agent: a system that facilitates LM agents to autonomously use computers to solve software engineering tasks. SWE-agent's custom agent-computer interface (ACI) significantly enhances an agent's ability to create and edit code files, navigate entire repositories, and execute tests and other programs. We evaluate SWE-agent on SWE-bench and HumanEvalFix, achieving state-of-the-art performance on both with a pass@1 rate of 12.5% and 87.7%, respectively, far exceeding the previous state-of-the-art achieved with non-interactive LMs. Finally, we provide insight on how the design of the ACI can impact agents' behavior and performance.

Subfields: Artificial Intelligence, Industrial and Manufacturing Engineering, Management Information Systems

View Paper: https://www.semanticscholar.org/paper/1c3c531fc0fbe79f97f367ed3648de8467caeeaa

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Chun Xia
2. Naman Jain
3. Zhi Jin
4. Kechi Zhang
5. Niklas Muennighoff
6. Terry Yue Zhuo
7. Wen-Ding Li
8. Jia Li

In Management Information Systems

1. Sam Dauncey
2. Adam Suma

Citations in the corpus

• BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions (2024, background): “We believe that there are three main expectations of a good execution-based programming benchmark: • The benchmark should be easy to use and efficient in evaluating the fundamental capabilities of LLMs. Repo-level benchmarks (e.g., SWE-bench [64]) are not suitable for this purpose.”
• AutoCodeRover: Autonomous Program Improvement (2024, neutral): “Figure 6 shows a visual summary of AutoCodeRover’s comparison with Swe-agent and Devin.”
• Neurosymbolic Program Synthesis (neutral): “Specifically, the emerging body of work on language-model agents for programming [17,18] and mathematical reasoning [19] do not just generate programs using a language model, but execute and analyze the resulting code and use the feedback from these processes to influence generation.”
• Agentless: Demystifying LLM-based Software Engineering Agents (2024, methodology): “To solve the challenging real-world software development problems from SWE-bench, inspired by the Devin AI Software Engineer [7], there has been a significant body of work from both academia and industry focusing on developing agent-based approaches [108, 37, 101, 30, 64, 29].”
• AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges (2025, neutral): “Programmatic pipelines automate this process using task templates, context fillers, and retrieval-augmented variables [232,233].”
• CodeAgent: Enhancing Code Generation with Tool-Integrated Agent Systems for Real-World Repo-level Coding Challenges (2024, background): “• Experimental results on nine LLMs demonstrate our C ODE A GENT ’s versatility and effectiveness in diverse code generation tasks, high-lighting its potential for real-world repository-level coding challenges.”
• SweRank: Software Issue Localization with Code Ranking (2025, unknown): “Effective localization is critical; without correctly identifying the relevant code segments, any subsequent attempt at automated repair is likely to fail or, worse, introduce new faults.”
• Neurosymbolic Repo-level Code Localization (2026, unknown): “Meanwhile, LLM-based retrieval techniques have been proposed to improve code localization performance by leveraging semantic understanding [3, 10, 17, 22, 26, 27, 29, 30, 32, 36].”
• Agent Retrieval Bench: Evaluating Repository Context Retrieval for Coding Agents (2026, unknown): “Modern agents may do this through a mixture of codebase indexes, repository maps, grep-like tools, file reads, and iterative search [1, 18], but the underlying question is the same: which repository context should enter the model before it edits?”
• SWE-Adept: An LLM-Based Agentic Framework for Deep Codebase Analysis and Structured Issue Resolution (2026, unknown): “We take an orthogonal perspective and employ working memory to store code-state checkpoints, facilitating reliable agent-driven version-control operations.”
• GraphLocator: Graph-Guided Causal Reasoning for Issue Localization (2025, unknown): “Despite its importance, automating this task is fundamentally difficult; even state-of-the-art large language models (LLMs) and frameworks [39, 45, 47] fail to localize the file to be changed over half of the issues in recent benchmarks [50].”
• Agentic Software Issue Resolution with Large Language Models: A Survey (2025, unknown): “Techniques Scaffold/MethodDesign End-to-EndScaffold Agent-Based MAGIS[147],AutoCodeRover[193],SWE-Agent[174],CodeR[26],MASAI[13],AlibabaLingmaAgent[107],OpenHandsCodeAct[157],CodexGraph[100],SpecRover[133],MarsCodeAgent[101],Hyper-Agent[126],SuperCoder[53],SWE-Search[12],InfantAgent[89],Learn-by-int”