Company AI agents fall short less because of weak models than because they lack knowledge of how a business works, LangChain co-founder and CEO Harrison Chase argued in a blog post published November 21. His proposed fix is what he calls a "knowledge factory," a system that gathers that knowledge and keeps it current for agents, with a dedicated "curator agent" doing the upkeep.
Chase wrote that LangChain runs its internal go-to-market and content studio agents on a knowledge factory. LangSmith Engine, the company's agent that reviews traces and suggests fixes, also maintains one for each agent it works on, according to the post.
Context versus knowledge
The post draws a line between the two. Context is "what an agent has available when doing a task," Chase wrote, while knowledge is "the reusable understanding we maintain so it can do that task correctly," such as what a metric means or which source to trust.
He tied the idea to software factories, where engineers build the systems and agents write the code. "General knowledge work is now following a similar path," Chase wrote, and he said that companies should put resources into building the knowledge agents need for those tasks.
Much of that knowledge is tacit, he wrote, sitting in employees' heads or spread across documents. He describes skills as procedures and scripts for doing tasks "the way your company expects," often never written down before.
Alongside skills and longer-lived domain knowledge about products and policies, Chase recommends maintaining what he calls data access guidance. If a company's data warehouse is already complete and current, "there's no need to maintain a separate copy," he wrote. The new work is keeping clear instructions on what data lives where and how to query it.
Flagging gaps, then curating
The factory starts with the agents doing the work. Chase recommends giving them a tool to report missing or wrong information they run into and, optionally, to suggest a correction.
Those agents should not make the fixes themselves, he argued. Asking agents "who are mainly trying to accomplish their own goals" to maintain shared knowledge "will cause it to become unruly and full of contradictions," the post says.
Instead, a curator agent runs in the background on a set cadence with access to raw sources such as Notion and Slack. It also reads agent traces to see how people use agents, along with the issues agents have flagged, according to Chase.
He called storage "the least important thing" in one sense, saying knowledge can sit as skill files in GitHub or as a graph in Neo4j, among other options that include LangChain's own LangSmith Context Hub. The industry has "converged on skills for 'procedural memory,'" he wrote, but no standard has emerged for semantic memory, and he named OKF as one possibility.
A worked example
The post walks through an agent preparing a revenue report from BigQuery and Salesforce data. The agent notices a renamed BigQuery column, works out the right one and flags the change.
A finance lead then corrects the report's formatting and asks for a new header that all LangChain documents now need. That night, the curator agent reviews the flag and the trace of that exchange and updates the relevant skills, according to the example.
Chase advised starting with "one recurring workflow," checking where agents struggle and closing that loop. The post includes no benchmarks or customer results. LangChain said it is working with a few customers to bring the approach to their agents.
Source: LangChain blog, Harrison Chase, “How to build a knowledge factory”
