Four researchers at Wavestone in France read through roughly 4 million lines of source code from 11 top AI coding tools, delivering a judgment that surprises practitioners: LangChain-style frameworks and vector retrieval (similarity-based search), which have been hot for three years, appear in zero production-grade tools. The 83-page paper racked up 1,329 likes and 107,000 views on X.

What this is

The study covers four first-party tools from major labs—Claude Code, Codex CLI, Gemini CLI, and Mistral Vibe—along with seven open-source projects including OpenHands and Aider. Together they comprise about 4 million lines of code written in three languages: Python, TypeScript, and Rust. The conclusion: every AI coding tool (a complete system where large models write code and run commands themselves, which academia calls coding agents) is composed of the same seven modules—the loop, tool invocation, context management, safety controls, the human interface, session storage, and extension interfaces. The differences are only in scale. Mini-SWE-Agent implements all seven modules in just 100 lines, while Codex uses 1.12 million lines of Rust. Yet their SWE-bench task completion rates are nearly tied.

Industry view

What sparked the most discussion are two "absences." First, not a single system uses mainstream agent frameworks like LangChain, AutoGen, or CrewAI—not even Google's own Gemini CLI uses Genkit. Every loop is hand-written (Python asyncio, TypeScript Promise, Rust Tokio), and tool registries are all custom-built on top of Pydantic or Zod. Second, not a single one uses vector retrieval to find code—instead they read code using ripgrep (a high-speed text search tool), tree-sitter (a code syntax parser), and the AGENTS.md project specification. The paper's explanation: code has deterministic structures like paths and syntax trees that semantic similarity cannot replace, and code changes every minute, making embeddings quickly outdated.

But there are dissenting voices. One view holds that coding agents cannot be directly generalized to customer service or knowledge-base scenarios; the Claude Code source analyzed in the paper is a March snapshot that may have missed later features. A deeper concern is that Codex has already started copying Claude Code's hook event names verbatim—the industry is rapidly converging, and the differentiation space left for latecomers is being compressed.

Impact on regular people

  • For enterprise IT: When vendors pitch RAG/Agent middleware, push back with one question—"Does your code actually use LangChain?"
  • For individual careers: Tools that use AI to write code will keep proliferating, but the moat lies elsewhere—in the ability to decompose tasks and write clear AGENTS.md files that help AI understand the project.
  • For the consumer market: Open-source coding agents will continue to converge; differentiation will ultimately come down to user interfaces and enterprise-grade security and compliance, while the underlying technology becomes more homogeneous.