Wink Pings

LangChain’s “Knowledge Factory”: Turning What Employees Never Wrote Down into AI Capabilities

LangChain co-founder Harrison Chase proposes the “knowledge factory” concept: when AI does work inside a company and gets things wrong, the gap is often knowledge—the ways the company operates that live in employees’ heads and have never been written down. He breaks the knowledge factory into three types of knowledge and three components.

Give AI longer context and stronger models, and it still gets things wrong inside your company. In a long essay, LangChain co-founder Harrison Chase argues that what’s missing is knowledge. He calls the system for collecting, organizing, and maintaining this knowledge the “knowledge factory.”

The best employees in a company carry a lot of tacit knowledge in their heads: how the company operates, what the operating standards are, where data lives, and how to access it. This knowledge is rarely written down. Sometimes it exists only in one person’s head; sometimes it is scattered across a pile of documents. The job of the knowledge factory is to turn it into a usable form and hand it to agents.

LangChain is already using this internally. Its go-to-market agents, content agents, and LangSmith Engine—which reviews run traces and suggests fixes—are all backed by a knowledge factory.

First, distinguish two terms. Context is the information an agent has at hand while doing a task. Knowledge is the understanding maintained over time so that a task can be done correctly: how a metric is calculated, which data source is trustworthy, and how to handle a certain situation. Some of it has been written down; more often it hasn’t. The job of the knowledge factory is to turn scattered documents and experience into context that future tasks can use.

Why call it a “factory”? Over the past year, the idea of the “software factory” has become popular. Rather than having the best engineers write software themselves, have them build systems that help everyone across the company write software more efficiently, with agents doing the actual development. Ordinary knowledge work is following the same path. Agents can already make slide decks, create videos, and do outbound calls. But to get them to do this work, you first have to feed them the company’s tacit knowledge. That’s where resources should go.

Knowledge comes in three types

He divides the knowledge to be maintained into three types.

The first is skills: processes, instructions, and scripts for doing things the way the company wants them done. Many have never been written down and are passed from colleague to colleague by word of mouth. They’re worth writing down.

The second is domain knowledge: core information that doesn’t go stale often—such as products, customers, competitors, and the organization. It has usually been written down somewhere, but it’s either incomplete or already outdated.

The third is data access guidance. Start with a good data strategy—Snowflake, Databricks, and the like. This is a prerequisite for agents to be able to do work. The data sources themselves should be complete and up to date; there’s no need to maintain another copy. What really needs to be maintained is the guide to “what data is where and how to query it.”

The factory has three parts

These three types of knowledge can’t be maintained manually; the workload is too large. He thinks a knowledge factory has three components.

The first is tooling that lets agents flag knowledge gaps. When an agent can’t find information while doing work, or discovers that existing information is wrong, it should report it. At minimum, it should report the problem; ideally, it should include a suggested correction.

The second is a curator agent, dedicated to updating and maintaining knowledge. You can’t rely on people for this, nor should you let the working agents make changes on the side. They’re good at spotting problems, but fixing them requires understanding the underlying knowledge and other data sources; letting them make the fixes will only make things messier. The curator agent runs periodically in the background. It can access the original sources where knowledge may be scattered—Notion, Slack, and the like—see how people are using agents, and then codify that experience into skills.

The third is knowledge storage. This is the least critical part—and also the least explored direction. Knowledge can be stored as skills in GitHub, as a file in Notion, in the OKF format in LangSmith Context Hub, or as a Neo4j knowledge graph. The curator agent needs to know how to work with these formats. The industry has already converged on skills for “procedural memory”; for more semantic memory, other standards may still emerge (OKF?). It’s too early to talk about this now, and it’s not worth over-engineering.

A closed loop from daytime to overnight

The article ends with a complete example.

An agent needs to prepare a revenue report. Base context: raw data lives in BigQuery and Salesforce. Knowledge: how to create reports in LangChain style is stored as a GitHub skill; how ARR is calculated is stored in the company wiki; how to access BigQuery and Salesforce is stored as a Context Hub skill.

The agent starts running and discovers that a column name in BigQuery has changed. It finds the correct column on the spot, flags the issue with the tool, and includes a suggestion. After the report is written, the finance lead and the agent make several formatting changes together and add a new requirement: all new LangChain documents must use a new header.

That evening, the curator agent sees the flag, reviews the run trace, notices this interaction, and updates the corresponding skills: use the new column for data retrieval and the new header for writing reports. The next day, the same problem no longer trips up the agent.

The conclusion is simple. Build the system first, let the work process expose knowledge gaps, fill those gaps, and do better next time. You don’t need to wait until everything is documented before letting agents start working. Start with one repetitive workflow, see where agents get stuck, and get the loop turning.

Original article: https://x.com/i/article/2124180107669766144

发布时间: 2026-11-22 03:44