SWE-agent: Agent-Computer Interfaces Enable…

SWE-agent asks how much software-agent capability comes from the language model and how much comes from the interface around it. Read it to see why summarized search, 100-line file windows, precise edits, lint feedback, and compressed history help one fixed model repair more repositories—and why a 12.47% resolve rate is mechanism evidence rather than production autonomy.

Reading focus: Why an agent-computer interface controls both the actions a model can express and the repository state it can observe. How the search, viewer, editor, linting, and history ablations isolate interface choices around a fixed model. How to read SWE-bench pass@1, API cost, sandboxing, patch quality, and manually tuned interfaces without turning a benchmark result into a deployment claim.

NeurIPS 2024. John Yang et al.. 35 min read, medium difficulty.