← Frameworks

SWE-agentSWE-agent

Princeton / Stanford（学术团队）Princeton / Stanford (academic teams)

开源Open source MITMIT 研究向Research-focused 代码修复Code fixes ACIACI SWE-benchSWE-bench partial lifecycle: active 核验Checked 2026-10-30

SWE-agent 是「agent 修代码」研究线的参考实现（MIT，Princeton/Stanford）： ACI（为 agent 设计的计算机界面）是它的核心命题，与 SWE-bench 评测生态同源。生产开箱不是它的定位——它是给研究与评测用的参考实现。

SWE-agent is a research harness for agent-driven code repair (MIT, Princeton/Stanford): ACI (agent-computer interface) is its core thesis, co-origin with the SWE-bench evaluation ecosystem. Production out-of-the-box is not its positioning — it is a reference implementation for research and evaluation.

适合与不适合When to use it, when not to

适合：评测/研究编码 agent、ACI 设计参考、SWE-bench 跑分。不适合：生产编码助手开箱；要产品形态。

Fits: evaluating/researching coding agents, ACI design references, SWE-bench runs. Not a fit: production coding assistants out of the box; product form factors.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-21

出品属地Origin: 国外出品foreign-made　Princeton / Stanford 学术团队（美国）。Princeton / Stanford academic teams (US).
国内使用Direct access: 大陆可直连direct OK　swe-agent.com 直连实测 200（2026-10-21）；GitHub 仓大陆直连波动（同日代理可达）。swe-agent.com measured 200 direct (2026-10-21); the GitHub repo's mainland reachability fluctuates (200 via proxy the same day).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-21 检索）；开源项目未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-21); the open-source project does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 各家 LLM 经接口层接入（OpenAI/Anthropic/HF 等，以官方文档为准，未逐条清点）；它的核心贡献是 ACI（agent-computer interface）——为 agent 设计的文件编辑/命令交互界面。
Various LLMs connect through an interface layer (OpenAI/Anthropic/HF etc., per the official docs, not itemized one by one); its core contribution is the ACI (agent-computer interface) — a file-editing/command-interaction interface designed for agents.
运行位置Where it runs: Python 库 + 命令行（pip 装）；GitHub 20,682★，最近推送 2026-10-26（API 核验）。学术出身（Princeton/Stanford 论文线）；SWE-bench 评测生态与它深度关联（同团队的基准线）。
A Python library + command line (pip install); GitHub 20,682★, last push 2026-10-26 (API verified). Academic origins (the Princeton/Stanford paper line); deeply tied to the SWE-bench evaluation ecosystem (the benchmark line from the same team).
本地文件Local files: agent 在容器或本机环境里读写文件、跑命令——文件操作真实发生在你给的环境；官方推荐容器化跑（隔离口径）。
The agent reads/writes files and runs commands inside a container or local environment — file operations really happen in the environment you give it; the official docs recommend running containerized (the isolation position).
关机后的任务Tasks after shutdown: 任务型运行：给定 issue→跑修复→结束；无常驻语义。批量评测跑（SWE-bench 式）是它的主用法之一。
Task-form runs: given an issue → run the fix → done; no resident semantics. Batch evaluation runs (SWE-bench style) are one of its main uses.
工具与扩展Tools & extensions: 工具面就是它的 ACI 设计核心：文件查看/编辑、搜索、命令执行—— 「为 agent 设计的界面」是它的研究命题，不是通用工具生态。
The tool surface is the core of its ACI design: file viewing/editing, search, command execution — "an interface designed for agents" is its research thesis, not a general tool ecosystem.
上下文与记忆Context & memory: 单任务上下文窗口管理是它的研究点（历史压缩/窗口策略）；跨会话记忆不是它的叙事。
Single-task context-window management is its research point (history compression / window policies); cross-session memory is not its narrative.
权限与限制Permissions & limits: 命令执行=真实权限面：agent 能跑 shell 命令，官方推荐容器隔离—— 裸机直跑的风险口径官方文档有提示（未逐条核验细节）。
Command execution = the real permissions surface: the agent can run shell commands, and the official docs recommend container isolation — the risk framing of running bare-metal is flagged in the official docs (details not itemized).
适合什么任务Best for: 适合：研究/评测「agent 修代码」能力（SWE-bench 线）、要 ACI 设计参考实现。不适合：生产编码助手开箱（它是研究向 harness）；要 GUI 产品形态。
Fits: researching/evaluating "agents fixing code" (the SWE-bench line), wanting a reference implementation of ACI design. Does not fit: a batteries-included production coding assistant (it is a research-oriented harness; productization is on you); wanting a GUI product form.

头号误解Top misconception

把它当生产编码 agent——它是研究向 harness（论文与评测线），产品化要自己补Treating it as a production coding agent — it is a research harness (paper and evaluation lines); productizing is on you
裸机直跑命令执行面——官方推荐容器隔离，别省这层Running the command-execution surface on bare metal — the official recommendation is container isolation; do not skip that layer
与 SWE-bench 基准混谈——SWE-bench 是评测集（同团队另一产物），SWE-agent 是跑评测的 agent 框架Blurring it with the SWE-bench benchmark — SWE-bench is the eval set (another product by the same team); SWE-agent is the framework that runs it

未知项清单Known unknowns

模型接口全目录（未逐条清点）
容器隔离的当前默认配置细节（未逐条核验）
与 mini-SWE-agent 轻量线的关系（官方有精简线，未逐条核验）

The full model-interface catalog (not itemized one by one)
Details of the current default container-isolation configuration (not itemized)
The relationship with the mini-SWE-agent lightweight line (the official docs have a slim line; not itemized)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-30，内容更新日content updated on 2026-10-21。

类型Type	名称Name	链接Link
repo	SWE-agent/SWE-agent · 仓库（MIT；20,682★，最近推送 2026-10-26，GitHub API 核验）SWE-agent/SWE-agent · repository (MIT; 20,682★, last push 2026-10-26, GitHub API verified)	https://github.com/SWE-agent/SWE-agent
docs	SWE-agent · 官方文档站SWE-agent · official docs site	https://swe-agent.com
changelog	SWE-agent/SWE-agent · Releases（changelog）SWE-agent/SWE-agent · Releases (changelog)	https://github.com/SWE-agent/SWE-agent/releases

实测记录Measured runs

本档未跑集成与性能实测；结论只覆盖 API 与官方文档披露的能力面。

This site has not completed measured runs. The test protocol is at protocols/framework/_protocol.md.