编程智能体技术参考资料

核心项目与文档

Claude Code

• 官方文档: https://docs.anthropic.com/en/docs/claude-code
• GitHub: https://github.com/anthropics/claude-code
• 核心架构: Agentic loop + tool use + context window management
• 关键特性: Subagent delegation, worktree isolation, hook system, MCP integration
• 参考论文: Anthropic "Building Effective Agents" (2024)

Cursor

• 官网: https://cursor.sh/
• 核心架构: Fork of VS Code + AI-first editing paradigm
• 关键特性: Tab completion (Copilot++), Cmd+K inline editing, Composer multi-file editing
• 技术基础: Tree-sitter AST, LSP integration, speculative decoding for completion

GitHub Copilot Workspace

• 官方博客: https://github.blog/product-coding/copilot-workspace/
• 核心架构: Plan → Code → Verify autonomous loop
• 关键特性: Task decomposition, step-by-step execution, environment sandboxing
• 技术基础: GitHub Actions runner, containerized execution

Aider

• GitHub: https://github.com/paul-gauthier/aider
• 核心架构: Chat-based coding with git integration
• 关键特性: /commands, repo map, multi-file editing, LLM switching
• 技术基础: Tree-sitter for repo map, git worktree for isolation

SWE-Agent

• GitHub: https://github.com/princeton-nlp/SWE-agent
• 论文: "SWE-agent: Agent-Computer Interactions Enable Autonomous Software Engineering" (2024)
• 核心架构: Agent-Computer Interface (ACI) design
• 关键特性: Custom action space, observation formatting, SWE-bench evaluation

OpenHands (原 OpenDevin)

• GitHub: https://github.com/All-Hands-AI/OpenHands
• 核心架构: Sandbox-based execution + LLM planning
• 关键特性: CodeAct agent, browser integration, runtime isolation

Devika

• GitHub: https://github.com/stitionai/devika
• 核心架构: Multi-agent planning + execution
• 关键特性: Project planning, step decomposition, web research

AutoCodeRover

• 论文: "AutoCodeRover: Autonomous Program Improvement via LLM" (2024)
• 核心架构: Spectrum-based fault localization + context retrieval
• 关键特性: Strata-based context, iterative repair

通用架构模式

• ReAct Loop: Reasoning + Action interleaving (Yao et al., 2023)
• Plan-then-Execute: Decomposition → sequential execution
• Tool Use: Function calling with structured I/O
• Context Management: Sliding window, RAG, compaction strategies
• Error Recovery: Retry, backtracking, human-in-the-loop escalation

评估基准

• SWE-bench: https://www.swebench.com/
• HumanEval: OpenAI code generation benchmark
• MBPP: Most Basic Python Programs
• LiveCodeBench: Real-time coding competition problems