跳到主要内容Skip to main content

编码 AgentCoding Agent

Coding Agent

  • 智能体Agents
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-17refs checked 2026-11-17
解决什么问题What problem it solves

能读代码库、改多个文件、跑测试并根据报错继续修的 AI 工具。An AI tool that reads a codebase, edits multiple files, runs tests and keeps fixing from failures.

这是什么What it is

与代码补全不同,编码 Agent 接受一个任务目标,自主规划并执行编辑、命令和验证。

SWE-bench 用真实 GitHub issue 检验这类系统能否产出通过测试的补丁。

Unlike code completion, a coding agent takes a task goal and autonomously plans edits, commands and verification.

SWE-bench tests such systems on real GitHub issues: can they produce a patch that passes the tests?

The coding-agent loop: read the issue and repository, edit across files, run tests, read failures and iterate.
编码 Agent 的循环:读 issue 与代码库,定位并修改多个文件,跑测试,读报错再修。The coding-agent loop: read the issue and repository, edit across files, run tests, read failures and iterate.
为什么需要知道Why it matters

它改变的是工程流程而不是单行输入速度;审查、权限和测试仍是必需,不能把「通过率」当成交付质量。It changes the engineering workflow, not just typing speed. Review, permissions and tests remain necessary; a pass rate is not delivery quality.

别名Chinese aliases

coding agent · 编码智能体 · 代码智能体

出处Sources
  • Jimenez 等 · SWE-bench(arXiv 2310.06770)Jimenez et al., SWE-bench: Can Language Models Resolve Real-World GitHub Issues? ↗
  • SWE-bench · 官方榜单SWE-bench leaderboard ↗