As of the end of 2025, LangChain has surpassed 110,000 stars on GitHub, making it the most widely used development framework at the AI application layer globally. This week we noted that it has bet the entire 1.x release line on a single function, create_agent()—it looks like a technical detail, but it reflects the industry's cognitive shift on "how AI applications should actually be written."
What This Is
LangChain is essentially an "AI application assembly framework." Its core value is not making models smarter, but assembling models, external tools, private data, and execution logic into a complete application.
In version 1.x, LangChain officially converges its main line to one formula: Agent = Model + Harness. The Model handles thinking and decision-making; the Harness manages Prompts, tools, state, middleware, and the execution loop.
To use an analogy: the model is the brain, the Harness is the body and the work rules. The Harness determines what tools the AI can use, whether it remembers things, whether sensitive operations require human approval, and what to do when an answer fails quality checks.
The create_agent() function introduced in 1.x is the official "minimum viable Harness" template—developers only need to fill in the model, the tool list, and middleware, and they can run an AI assistant that calls external systems.
Industry View
Supporters argue that LangChain solves a real pain point: different model vendors have different API formats and tool-calling conventions, so switching models often means rewriting half the business code. Models themselves cannot access enterprise private data, query real-time databases, or complete multi-step tasks—and the Agent loop (letting the model decide whether to call a tool, which one to call, and when to stop) is precisely the standard approach to filling this gap.
But we also need to acknowledge the opposition. First, LangChain's abstraction layer adds learning costs; for beginners who have only used the OpenAI API, there is a significant conceptual gap between "calling a chat endpoint" and "building a Harness." Second, the framework is still iterating rapidly, with breaking changes between 1.x and older versions—enterprises deploying it in production should be mentally prepared for upgrade pace. Third, there are voices in the industry questioning whether Agent's real-world effectiveness is often overstated, and whether the latency, cost, and error rate brought by multi-step tool calls are really worth it in serious business scenarios.
Impact on Regular People
For enterprise IT: Over the past year, more and more enterprises have begun using "can we build an internal AI assistant with LangChain" as a litmus test for evaluating AI capability. If your team is assessing AI implementation projects, understanding the basic logic of the Agent framework will become a baseline requirement.
For individual careers: Even if you don't write code, understanding the "Model + Harness" line of thinking helps you evaluate vendor proposals—when a company pitches you an AI product, you can ask: how does its Harness handle memory, permissions, and audit? This is often more critical than the model itself.
For the consumer market: Frameworks like LangChain lower the development barrier for AI applications, meaning more vertical AI products will emerge in the coming years—custom assistants for law firms, clinics, and e-commerce sellers, for example. They will be priced higher than general-purpose chatbots, but far cheaper than hiring a human.