此语言仍有 24 条内容或界面任务待翻译;未完成部分显示中文原文。
AI Tools Navigation / Agents & Automation / GitHub Copilot 云代理(Copilot Cloud Agent)
智能体与自动化 · GitHub / Microsoft

GitHub Copilot 云代理(Copilot Cloud Agent)

Dispatch tasks on GitHub in natural language and have an agent asynchronously generate code changes and submit pull requests for human review.

AI 辅助整理
Code GenerationPull RequestsCode MaintenanceTest CoverageDocumentation

Tool Introduction

The GitHub Copilot cloud agent is an autonomous asynchronous development agent integrated into GitHub: it receives tasks in the form of natural language, code snippets, or images from issues, PR comments, or Copilot Chat, autonomously creates branches, explores the codebase, generates implementation plans and draft code, and can open pull requests, supporting use cases such as bug fixes, incremental features, documentation, tests, and prototyping. The agent modifies code, runs tests and linters in an isolated ephemeral development environment, with an anti-data-exfiltration firewall enabled by default, and pushes are restricted to copilot/ prefixed branches. The resulting code must be verified through human review and cannot be treated as a fully automatic replacement.

The workflow consists of four steps: first, the task and context are combined into a prompt via an issue, comment, or chat message; next, a large language model analyzes the input and invokes the necessary tools; then natural language and code suggestions are generated; finally, the pull request description is updated after the first run completes. Users can comment on the PR or @copilot to provide feedback, and the agent will resubmit that feedback to the model to continue iterating.

Ways to improve performance: keep the task scope clear—state the problem to solve, acceptance criteria (such as whether unit tests are needed), and files that may need changes; add custom Copilot instructions so the agent knows how to build, test, and verify; you can also customize the development environment, configure the firewall, and extend capabilities via MCP (Model Context Protocol).

Security measures include: the agent only responds to users with write access, and workflows triggered by agent PRs require approval from a user with write access; hidden characters are filtered to reduce prompt injection risk; the agent can only access repositories it is scoped to and cannot access other repositories; at runtime it has no access to organization/repository Actions secrets and variables—only secrets/variables added in the Copilot environment are passed in; a firewall is enabled by default to prevent exfiltration of code or sensitive data.

Limitations: quality varies depending on the codebase, input, and language coverage; training data and retrieved context may carry bias; generated code may be syntactically correct but insecure or misaligned with intent, requiring thorough review; even with blocking in place, output may still match or closely match public code and may not provide attribution; users are responsible for ensuring legal compliance and avoiding uses prohibited by the terms of service.

Who It's For

Engineers and teams developing with GitHub, especially maintainers who need to handle maintenance tasks such as dependency upgrades, security patches, incremental feature development, documentation updates, and test additions; outputs require human review and are not suitable for scenarios that bypass code review and merge directly.

Recommended Steps

  1. 01

    On GitHub.com, describe the task to the agent in natural language via an issue, pull request comment, or Copilot Chat, attaching acceptance criteria and hints about relevant files

  2. 02

    Wait for the agent to create a branch, explore the codebase, and generate an implementation plan and draft code; after the first run completes, review its updated pull request description

  3. 03

    Comment on the pull request or @copilot with feedback so the agent iterates on the changes based on that feedback

  4. 04

    Review and test the agent-generated code, performing secure coding checks, IP scans, and vulnerability checks before merging

  5. 05

    If you run into problems, provide feedback via the thumbs-up/thumbs-down icons below the response or through the community discussion forum

Know Before You Use

The agent only responds to users with write access; workflows triggered by agent PRs require approval from a user with write access before they can run. Pushes are restricted to copilot/ prefixed branches and cannot be pushed to the default branch; at runtime the agent has no access to organization/repository Actions secrets and variables—only secrets/variables configured in the Copilot environment are passed in. English is the primary supported language, and performance varies by codebase and input. See the official website for specific subscription plans.

费用与套餐:以官网当前说明为准,本站不提供套餐价格保证。

Product Selection Assistant

让需求变成可修改的筛选条件

依据已录入参数生成条件,结果可查看和修改。帧率与性能需核对实测,不能仅凭算力数值保证。