“SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering,” posted to arXiv on May 6, 2024 by John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press - the Princeton group behind SWE-bench - made an argument that shifted how people built coding agents. The bottleneck, they claimed, was not only the model but the interface through which the model touches the computer. So they designed an Agent-Computer Interface, or ACI: a custom set of commands tailored to how a language model best works, for opening and editing files, navigating a whole repository, and running tests.
A good ACI is not the same as a good human interface. The paper showed that small design choices - how file contents are shown, how edits are applied and linted, how search results are formatted - make a large difference to how reliably the agent operates. With its ACI, SWE-agent resolved 12.5 percent of issues on SWE-bench, far above the previous best from non-interactive language models at the time, and 87.7 percent on HumanEvalFix.
SWE-agent was released as open source and became a widely used foundation and reference point for autonomous coding agents. Its central lesson - that designing the agent’s interface to the world is a first-class engineering problem, on par with choosing the model - carried directly into the agentic coding tools that followed.