AI Coding Guide

English

简体中文

English

简体中文

Appearance

Appendix B: Technology Radar ​

This technology radar helps you understand technology maturity and adoption recommendations in the AI programming field. Divided into four quadrants:

  • Adopt: Production-ready, widely adopted, worth using immediately
  • Trial: Worth trying, high potential, can trial in non-critical projects
  • Assess: Interesting but needs observation, wait for maturity or standardization
  • Hold: Avoid using or gradually phase out

Adopt ​

Models ​

  • GPT-4o: OpenAI's latest multimodal model, excellent comprehensive performance, stable API
  • Claude Sonnet 4.6: Exceptional programming task performance, 200K context window
  • DeepSeek-V3: Open-source alternative, extremely cost-effective, locally deployable

Tools ​

  • Cursor: Most mature AI programming tool, complete VSCode ecosystem
  • Claude Code: Terminal-based AI assistant, suitable for scripts and DevOps
  • GitHub Copilot: Code completion standard, good team collaboration support

Frameworks ​

  • OpenAI SDK: Official SDK, comprehensive documentation, active community
  • Vercel AI SDK: Best practices for streaming UI, convenient React integration
  • LangChain: Most complete RAG and Agent ecosystem

Protocols ​

  • OpenAI API: De facto standard, best compatibility
  • REST/SSE: Mature streaming communication solution

Databases ​

  • PostgreSQL + pgvector: Vector search capability, production-grade stability
  • Redis: First choice for caching and session management

Platforms ​

  • Vercel: Best choice for Next.js deployment
  • Cloudflare Workers: Edge computing, global low latency
  • Railway: Rapid deployment, developer-friendly

Trial ​

Models ​

  • Gemini 2.5 Flash: Google's latest model, strong multimodal capability, cheap price
  • GPT-4o mini: Cost-effective alternative, suitable for batch tasks
  • o1: Strong reasoning capability, suitable for complex logic tasks (but higher cost)

Tools ​

  • Windsurf: Codeium's newly launched AI IDE, innovative features
  • Zed: High-performance editor, gradually improving AI integration
  • Aider: Command-line AI programming tool, suitable for automation

Frameworks ​

  • LangGraph: LangChain's Agent framework, suitable for complex workflows
  • CrewAI: Multi-agent collaboration framework, concise API
  • AutoGen: Microsoft's Agent framework, high research value

Protocols ​

  • MCP (Model Context Protocol): Anthropic's context standard, rapidly developing ecosystem
  • OpenAI Realtime API: Voice real-time interaction, low latency

Databases ​

  • Supabase: Postgres managed service, integrated vector search
  • Weaviate: Professional vector database, excellent performance

Platforms ​

  • Replit: Online development environment, tightly integrated AI assistant
  • Modal: Serverless GPU, suitable for model inference

Assess ​

Models ​

  • Llama 4: Meta's open-source model, awaiting release
  • Gemini 2.5: Google's next-generation model, unclear roadmap
  • Domestic Closed-Source Models: Tongyi Qianwen, Wenxin Yiyan etc., API stability needs improvement

Tools ​

  • Claude Desktop: Anthropic's desktop app, limited features
  • Copilot Workspace: GitHub's AI IDE, early stage
  • Devin: AI software engineer, restricted access

Frameworks ​

  • OpenAI Agents SDK: Newly released Agent framework, incomplete documentation
  • Semantic Kernel: Microsoft's AI orchestration framework, immature ecosystem
  • Haystack: RAG framework, intense competition

Protocols ​

  • A2A (Agent-to-Agent): Google's proposed Agent protocol, under standardization
  • ANP (Agent Network Protocol): OpenAI proposal, spec not public

Databases ​

  • Pinecone: Professional vector database, but expensive
  • Chroma: Lightweight vector database, relatively simple features
  • Qdrant: Vector database newcomer, ecosystem needs improvement

Platforms ​

  • Val Town: Online programming platform, small community
  • E2B: Agent sandbox environment, limited use cases

Hold ​

Models ​

  • GPT-3.5: Already replaced by GPT-4o mini, outdated performance
  • Old Claude: Claude 3 Opus high cost, 3.5 Sonnet better

Tools ​

  • Early AI IDEs: Niche tools with incomplete features, unstable maintenance
  • Pure Prompt Tools: ChatGPT web version for programming, low efficiency

Frameworks ​

  • Outdated LLM Frameworks: Like GPT-Index (renamed LlamaIndex), chaotic ecosystem
  • Complex Agent Frameworks: Over-designed, high learning cost, low actual benefit

Protocols ​

  • Custom Protocols: Non-standardized proprietary protocols, poor compatibility
  • GraphQL for LLM: Over-designed, REST + SSE sufficient

Databases ​

  • Traditional Relational DBs for Vectors: Like MySQL, insufficient performance
  • Self-built Vector Index: Unless special needs, use mature solutions

Platforms ​

  • Old Platforms without AI Support: Traditional PaaS like Heroku, lack AI tool integration
  • Blockchain AI Platforms: Concept over practicality, high cost

Selection Recommendations ​

Rapid Development ​

  • Models: GPT-4o or Claude Sonnet 4.6
  • Tools: Cursor + Claude Code
  • Frameworks: Vercel AI SDK + OpenAI SDK
  • Deployment: Vercel or Cloudflare Workers

Cost-Sensitive ​

  • Models: GPT-4o mini or DeepSeek-V3
  • Tools: GitHub Copilot (enterprise already purchased)
  • Frameworks: Direct API calls, reduce middleware
  • Deployment: Railway or self-hosted servers

Enterprise Scenarios ​

  • Models: Multi-model combination (main GPT-4o, backup Claude)
  • Tools: Cursor (dev) + Claude Code (ops)
  • Frameworks: LangChain (rich ecosystem)
  • Deployment: Private cloud + API gateway

Research Exploration ​

  • Models: o1 (reasoning) + Gemini 2.0 (multimodal)
  • Tools: Windsurf or Zed (experience new features)
  • Frameworks: AutoGen or LangGraph (experiment Agent)
  • Deployment: Modal (GPU inference)

Last updated: 2026-02-20

An AI coding guide for IT teams

MIT License