编码智能体辅助软件工程

形态

16x

16x Prompt 示例：

Webwright : Shell 增强

graph TD
    A[User] -->|Enters command| B[Webwright Shell]
    B -->|Processes command| E{OpenAI or Anthropic?}
    E -->|OpenAI| F[OpenAI API]
    E -->|Anthropic| G[Anthropic API]
    F -->|Response| H[Process AI Response]
    G -->|Response| H
    H <-->|Query/Update| L[(Vector Store)]
    H <-->|Query/Update| M[(Set Store)]
    H -->|Generate Code/App| I[Code/Application Output]
    H -->|Execute Function| J[Function Execution]
    J -->|Result| K[Process Function Result]
    K -->|Update Context| B
    I -->|Display to User| A
    B <-->|API Calls| N[mitta.ai API]
    N -->|Document Processing| O[Process Documents]
    N -->|Web Crawling| P[Crawl Websites]
    N -->|Other Functionality| Q[...]

Auto Wiki

https://wiki.mutable.ai/

Auto Wiki v2

CrawAI

https://github.com/crewAIInc/crewAI

• 基于角色的智能体设计：可定制具有特定角色、目标和工具的智能体。
• 自主智能体间委托：智能体能够自主地委托任务并在彼此间进行询问，从而提高问题解决效率。
• 灵活的任务管理：定义任务并使用可定制工具，动态分配给智能体。
• 流程驱动：目前支持顺序任务执行和分层流程，正在开发更复杂的流程，如共识和自主流程。
• 将输出保存为文件：可以保存单个任务的输出为文件，以便后续使用。
• 输出解析为 Pydantic 或 JSON：如需，可将单个任务的输出解析为 Pydantic 模型或 JSON。
• 兼容开源模型：可使用 Open AI 或开源模型运行您的团队。参考“连接 crewAI 到 LLMs”页面，了解如何配置智能体与模型之间的连接，甚至包括本地运行的模型！

示例

Builder.io

Introducing Micro Agent: An (Actually Reliable) AI Coding Agent

Micro Agent 是一种生成式 AI 工具，通过聚焦于特定任务并通过单元测试提供清晰的确定性反馈来创建代码。以下是其工作流程的细分：

1. 描述你的函数：你提供一个关于所需函数的自然语言描述。
2. AI 生成测试：基于你的描述，Micro Agent 创建定义函数预期行为的单元测试，包括多个输入输出场景。
3. AI 编写代码：使用大型语言模型（LLMs）的能力，Micro Agent 编写代码（支持 JavaScript、TypeScript、Python 或其他语言）以通过生成的测试。
4. 自动迭代：如果初始代码未能通过任何测试，Micro Agent 会反复编辑源代码并重新运行测试，直到所有测试都通过。这确保了最终代码符合指定要求。

主要优势

• 更高的保证：生成的代码更可靠，能够满足指定的要求，因为它通过了确定性的测试。
• 简化的开发过程：自动化的迭代过程有助于高效地创建高质量代码，节省时间并提高结果的信心。

Codium: Cover-Agent && PR-Agent

CodiumAI Cover Agent aims to help efficiently increasing code coverage, by automatically generating qualified tests to enhance existing test suites.

Below is the roadmap of planned features, with the current implementation status:

• Automatically generates unit tests for your software projects, utilizing advanced AI models to ensure comprehensive test coverage and quality assurance. (similar to Meta)
  • Being able to generate tests for different programming languages
  • Being able to deal with a large variety of testing scenarios
  • Generate a behavior analysis for the code under test, and generate tests accordingly
  • Check test flakiness, e.g. by running 5 times as suggested by TestGen-LLM
• Cover more test generation pains
  • Generate new tests that are focused on the PR changeset
  • Run over an entire repo/code-base and attempt to enhance all existing test suites
• Improve usability
  • Connectors for GitHub Actions, Jenkins, CircleCI, Travis CI, and more
  • Integrate into databases, APIs, OpenTelemetry and other sources of data to extract relevant i/o for the test generation
  • Add a setting file

Factory.ai：化繁为简

软件开发包含一系列的任务及子任务，诸如于：

• 依赖编码规范、框架知识、架构设计、安全指标等进行代码编写

如何结合 AI 更好进行知识管理，简化为项目、编码、测试、检视、文档等步骤，实现从想法到产品的快速迭代？

Cursor：Shadow Workspace

Shadow Workspace: Iterating on Code in the Background

Shadow Workspace 设计标准。我们希望 Shadow Workspace 能实现以下目标：

1. LSP 可用性：AI 应该能够看到他们更改后的 lint 提示，能够跳转到定义，并且更广泛地与语言服务器协议 (LSP) 的所有部分进行交互。
2. 可运行性：AI 应该能够运行他们的代码并查看输出。

我们最初专注于 LSP 可用性。 这些目标应满足以下要求：

1. 独立性：用户的编码体验必须不受影响。
2. 隐私性：用户的代码应是安全的（例如，所有代码都保存在本地）。
3. 并发性：多个 AI 应该能够同时进行工作。
4. 通用性：应适用于所有语言和所有工作区设置。
5. 可维护性：代码应尽可能少且易于隔离。
6. 速度：任何地方都不应有分钟级的延迟，并且应有足够的吞吐量以支持数百个 AI 分支。

其中许多反映了为超过十万用户构建代码编辑器的现实情况，我们确实不希望对任何人的编码体验产生负面影响。

问题：

实现：

1. AI 提议对文件进行编辑。
2. 编辑从普通窗口的渲染进程发送到其扩展宿主，然后传送到 Shadow window 的扩展宿主，最后传送到 Shadow window 的渲染进程。
3. 编辑在 Shadow window 中应用， Shadow window是隐藏且独立于用户的，所有 lint 提示都以相同的方式发送回来。
4. AI 接收到 lint 提示，并决定如何进行迭代。

shadow_workspace.proto

https://gist.github.com/arvid220u/b976c87c7ec9f6f66595dc0ebc0f07d6

syntax = "proto3";

package aiserver.v1;
option go_package = "github.com/anysphere/everysphere/schema/aiserver/v1";

service ShadowWorkspaceService {
	rpc GetLintsForChange(GetLintsForChangeRequest) returns (GetLintsForChangeResponse) {}
  rpc ShadowHealthCheck(ShadowHealthCheckRequest) returns (ShadowHealthCheckResponse) {}
}

message ShadowHealthCheckRequest {}

message ShadowHealthCheckResponse {}

// only lints in the supplied files are returned
//
// note that this only returns lints that were changed! unless `get_all_lints_not_just_delta_lints_for_range_in_final_model` is set
message GetLintsForChangeRequest {
  repeated File files = 1;

  message File {
    string relative_workspace_path = 1;

    // the "change" we are getting lints for is the implicit delta between `initial_content` and `final_content`
    string initial_content = 2;
    string final_content = 3;

    // for any lints in the final model that are inside one of these ranges, we return it regardless of whether it existed in the initial model too
    optional RangeCollection get_all_lints_not_just_delta_lints_for_ranges_in_final_model = 4;

    message RangeCollection {
      repeated IRange ranges = 1;
    }

    // simply corresponds to vscode's irange
    message IRange {
      // 1-indexed
      int32 start_line_number = 1;
      // 1-indexed
      int32 start_column = 2;
      // 1-indexed
      // inclusive
      int32 end_line_number = 3;
      // 1-indexed
      int32 end_column = 4;
    }
  }

  // quick fixes will be returned for each lint. potentially a little bit slower
  bool include_quick_fixes = 2;

  // if a new file is not created, some tools will report inaccurate or incomplete lints (e.g., the typescript language server won't properly detect which tsconfig is the right one)
  // the temporarily created file will have a .shadowworkspace-uuid.ts extension, to reduce the chance of conflicts
  // the hope is that it won't affect the user, but it may, so therefore this is off by default for now
  // once we have a proper proxy folder structure, then this should hopefully be obsolete
  // WARNING: this can cause problems for the user!!! (our yarn watch breaks when new files are added and deleted, for example)
  // do not run this for real users! before we have the proxy folder set up
  bool do_not_use_in_prod_new_files_should_be_temporarily_created_for_increased_accuracy = 3;

}

message GetLintsForChangeResponse {
  repeated Lint lints = 1;
  message Lint {
    string message = 1;
    string severity = 2;
    string relative_workspace_path = 3;
    // the position refers to the position in the `final_content` model
    int32 start_line_number_one_indexed = 4;
    int32 start_column_one_indexed = 5;
    int32 end_line_number_inclusive_one_indexed = 6;
    int32 end_column_one_indexed = 7;

    message QuickFix {
      message Edit {
        string relative_workspace_path = 1;
        string text = 2;
        int32 start_line_number_one_indexed = 3;
        int32 start_column_one_indexed = 4;
        int32 end_line_number_inclusive_one_indexed = 5;
        int32 end_column_one_indexed = 6;
      }
      string message = 1;
      string kind = 2;
      bool is_preferred = 3;
      // TODO: this edit can also be a file edit in vscode! currently, we ignore those, and only include text edits
      repeated Edit edits = 4;
    }
    // only included if `include_quick_fixes` is true
    repeated QuickFix quick_fixes = 9;
  }
}

其它

相关材料

开源自主 AI Agent 汇总 伴随 LLM 的快速发展，以 LLM 为大脑的 Agents 能力也越来越强，它们能结合规划、记忆和工具使用，来让 LLM 具备丰富的行动能力。

关于 Autonomous Agents LLM Powered Autonomous Agents via Lilian Weng https://lilianweng.github.io/posts/2023-06-23-agent/… 今天我们一起看看都有哪些开源 Agent 项目值得参考 以下是整理好的表格：