我对 AI 编码代理领域的未经过滤的看法

Agentic Coding Overview: It's the hottest tech space with many companies involved. There's a mix of hype and real differences. Understanding what agents can do is like an "art" and requires hands-on experience.
Some Background: The author believes in the future of agentic coding for large-scale projects. Their day job is on an open-source Rust project. Agents are still expensive and there are many tools, making it hard to choose.
Specific Product Analysis:
- Cursor: The ambitious frontrunner with features like the TAB feature. Its pricing and version changes show its direction. It's going head-to-head with Devin and aims to be a comprehensive platform. Its moat is speed.
- VS Code/GitHub Copilot: Copilot was a milestone but has been surpassed. VS Code is likely to reclaim the throne as AI coding becomes a consensus.
- Claude Code: A CLI-based agent with detailed prompts. It's moving forward with features like Claude Code 1.0 and availability to more users. Anthropic's intention with it is unclear but likely related to data and model training.
- Amp: Has a unique product philosophy and a leaderboard feature. It's not yet profitable but has potential.
- OpenAI Codex (in ChatGPT): OpenAI's fully automatic coding agent in ChatGPT aims to make ChatGPT a dispatching hub. It has a superior overall experience.
- Devin: Started expensive but now has a pay-as-you-go model. It has integrations and a polished product but faces competition.
- v0: Focuses on front-end UI prototyping with React-based generation. It's releasing its own model and integrating with existing codebases.
- Bolt / Replit / Lovable: "Idea to App" platforms for end-to-end development. They have a smooth vibe coding experience but face limitations in serious scenarios.
- YouWare: A platform for user-generated software, hiding code from non-coders. It has addictive features but faces challenges in monetization and user expectations.
Big Picture: Industry Landscape & Technical Directions:
- Market Segmentation: AI coding includes AI-assisted coding, end-to-end agents, and vibe coding/UGS.
- The Awkward "Half-Baked" State: Agents are still not perfect but are improving. The cost is a barrier, and there's a trade-off between performance and cost.
- What Capabilities Does an Agent Need: Memory, long context capability, task management, and proactive communication are important. Building a good coding agent may require being a good user.
Final Thoughts: Choosing tools depends on understanding "craft" and prompt engineering. Agents will likely get better with use. The development of AI raises questions about what to generate with infinite power and if studying it is worth it. Products like YouWare might offer an answer.