We noticed a recent 600+-line C++ source-code walkthrough on Juejin: engineer chen_dl has open-sourced his self-built AI Agent server codebase on Gitee, with main.cpp wiring together the DeepSeek LLM, MCP (the standard protocol letting models call external tools), and A2A (the spec for AI agents to talk to each other) in an industrial-grade bundle. What matters here—it signals Chinese developers are now seriously tackling the AI Agent "plumbing" layer.
What This Is
This project isn't about algorithms—it's all wiring. main.cpp cuts the startup flow into seven stages: read env vars and command line, load five configuration segments (DeepSeek, logging, MCP, compression, project instructions), construct shared components, inject AgentRuntime, register callbacks, and finally spin up two network services—one running A2A synchronous messaging, one running the coding agent's Web streaming interface.
Trade-offs are written into the entry file: missing DeepSeek config triggers an immediate crash exit; MCP load failure degrades gracefully and continues. These decisions are scattered across a dozen callbacks. 607 lines of main.cpp is the engineering reality—it's also a working sample of what an Agent service actually looks like.
Industry View
The positive signal we see: Chinese developers are now getting their hands dirty on Agent infrastructure. DeepSeek ships the model; MCP and A2A ship the connection standards. Previously, "how to glue them together" relied mostly on Silicon Valley open source. Now Chinese developers are writing it themselves—and willing to break every line down for readers.
But cold water needed: this is a personal project, far from enterprise-ready. main.cpp is already 600+ lines, and module boundaries are starting to blur. More critically, MCP is still evolving fast (Anthropic is still iterating), and A2A lacks true convergence—every vendor pushes its own variant. Code based on any given version may need substantial rewrites when the protocols update.
Our judgment: the direction is right, the engineering effort is real, but it's still early for true infrastructure-grade plumbing.
Impact on Regular People
For enterprise IT: When evaluating AI Agent products, use this open-source repo as a reference—see what a minimum viable version requires, and compare against what vendor solutions are missing.
For individual careers: Engineering roles are stratifying. People who "understand protocols and can wire different AI systems together" will become increasingly valuable; terms like MCP and A2A will show up in more job descriptions.
For consumer markets: No direct short-term impact. But once standards like MCP achieve real adoption, different AI tools will be able to interoperate—and only then will today's mess of "each app has its own AI, none talking to each other" be resolved.