Stop oversimplifying Coding Agent as just a 'wrapper' for LLM. That’s like saying TSMC is just a 'wrapper' for ASML. If you only have lithography machines but no manufacturing processes, you’ll never produce chips with acceptable yield. When you apply this logic to AI development, the division of labor between the two becomes crystal clear. LLM: Equivalent to ASML’s lithography machines (core equipment) LLM is the core engine for understanding natural language and code. Just like EUV lithography machines, it’s an indispensable foundational capability. Currently, there are only a handful of top-tier model providers globally (OpenAI, Anthropic, etc.), offering general-purpose intelligence and reasoning capabilities. Coding Agent: Equivalent to TSMC’s fabs (manufacturing processes) Having top-notch equipment doesn’t mean you can instantly produce 3nm chips. Similarly, real software development is far from 'linear code generation'; it’s a systematic engineering process. The core strength of Coding Agent lies in building an 'advanced process' to harness LLM: - Context management: akin to refining and precisely dispatching raw materials (information) - Task planning and reasoning: comparable to orchestrating intricate and complex process logic - Automated error correction: like real-time yield detection and repair during production Of course, this is just the tip of the iceberg. Beneath the surface, real engineering systems involve hundreds or even thousands of hidden fine-tuning techniques and specialized workflows. This is why the same model (lithography machine) performs vastly differently in the hands of Cursor or Claude Code compared to a simple script. It’s also why latecomers like Japan’s Rapidus, even after acquiring the same EUV lithography machines, can’t replicate TSMC’s 3nm process overnight. Because this involves a wealth of unquantifiable, hard-to-standardize tacit knowledge and engineering optimization experience. Coding Agent is far from a simple API wrapper. It’s a production system with complex orchestration logic, striving for 'optimal yield.' In this system, how to customize workflows based on project characteristics and how to maximize the potential of the model will be the main battleground for future differentiated competition.