AIToday
Large Language ModelsAI Coding AssistantsLangChain BlogPublished: Nov 22, 2026, 04:00 JST

LangChain lays out a "knowledge factory" for AI agents

Send on LINE
LangChain lays out a "knowledge factory" for AI agents

3 Key Points

  1. What happened

    LangChain described a "knowledge factory" for AI agents: agents flag missing knowledge, a background curator agent updates it, and it is stored in formats like GitHub skills or LangSmith Context Hub.

  2. Why it matters

    The post argues agents fail on missing company knowledge, not weak models, so the maintenance loop may matter more than the storage format, which it says is the least important piece.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The post frames the problem less as model quality and more as missing institutional knowledge — the tacit understanding that the most efficient employees carry in their heads, about who to talk to and why things are done a certain way. It distinguishes this knowledge from context: context is what an agent has available when doing a task, while knowledge is the reusable understanding kept so the task is done correctly, such as what a metric means or which source to trust.

The structural argument borrows from the "software factory" idea, in which the best engineers build systems that let everyone else build software. LangChain applies the same logic to general knowledge work: presentations, launch videos and outbound tasks can be handed to agents, but only if the scattered knowledge behind them is collected and maintained. That is why the post says resources should go into the factory rather than into documenting everything up front; the stated goal is a loop where doing the work reveals what knowledge is missing and closing those gaps makes future work better.

The architecture is deliberately light on storage. For procedural memory the post says the industry has converged on skills, while for semantic memory it expects a different standard may emerge and says it is too early to be prescriptive. The heavier lift sits with the curator agent, which runs on a cadence with access to raw sources like Notion and Slack, to agent traces, and to the issues agents flag — in the worked example, it notices a renamed BigQuery column and a finance lead's formatting fixes, then updates the relevant skills.

FAQ
What are the three kinds of knowledge LangChain says to maintain?
Skills (procedures and scripts for doing work the company's way), domain knowledge (definitions, policies and facts about products, customers and competitors), and data access guidance (which sources to use and how to query them).
Why not just have the working agents update the knowledge themselves?
The post says agents are good at flagging what is wrong or missing, but fixing it needs deep understanding of the underlying knowledge and other data sources. Letting them maintain it directly would make it unruly and full of contradictions.
Where is the knowledge stored?
The post says storage is the least important piece: a skill format in GitHub, a single file in Notion, OKF format in LangSmith Context Hub, or a knowledge graph in Neo4j all work.
LangChain BlogRead Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleShortical debuts AI series "Inevitable" with Aki Avni