SWE-agent¶

An Agent–Computer Interface for browsing, editing, and testing code repositories with software-engineering agents.

Paper information¶

Field	Details
Paper	SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
Organization	Princeton University
First public release	2024-05-06
Authors' code	Public implementation
Adapter / method	`swe-agent`
Local reproduction code	`src/auto_research/agent_research/`

Original paper¶

Motivation and method¶

A general-purpose shell exposes an overly broad action space and verbose output to an LLM. SWE-agent's specialized ACI constrains repository search, file viewing, precise editing, and testing so the model can locate the cause of an issue and validate a patch.

flowchart LR
    I["GitHub issue"] --> L["Find source and tests"]
    L --> E["Edit precisely"]
    E --> T["Run regression tests"]
    T -->|Fail| L
    T -->|Pass| P["Patch"]

Key figure from the paper¶

Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ a_t\sim\pi(\cdot|o_{\le t},I),\qquad o_{t+1}=\operatorname{ACI}(R_t,a_t). \]

Paper results¶

SWE-bench shows benefits from ACI when fixing real GitHub issues. It is an offline software-engineering benchmark, not a production online A/B test.

Local reproduction¶

The local ACI exposes only file reads, writes to solution.py, and a fixed unittest command. Each starting repository must fail a real test; the agent locates the issue, edits, and validates by exit code.

auto-research agent-eval --method swe-agent --benchmark swebench-local \
  --episodes 12 --seed 42

See the stable metrics.

Reproduction boundaries¶

This is an executable SWE-style microbenchmark, not official SWE-bench Lite data or its Docker images. The swebench-local name preserves that distinction. Local success is not the paper's resolved rate.