LangChain founder Harrison Chase recently published the article "How to build a knowledge factory," arguing that the key bottleneck for enterprise AI adoption is not better models or more context, but "better knowledge" — how a company operates, its operating standards, where data lives, and how to access it. The view sparked a chain of discussion, including both hands-on practice sharing and developers' reflection on the relationship between AI and humans.
Core argument: knowledge matters more than models
In the article, Chase analyzes that the most efficient employees carry knowledge in their heads: they know who to ask, where to find answers, and why things are done this way. This knowledge is often tacit and scattered.
For AI to truly work in the enterprise, this knowledge needs to be collected, organized, and continuously maintained. Based on this idea, Chase proposed the concept of a "knowledge factory," stressing that enterprises need to systematically build knowledge management processes rather than relying solely on model capability improvements.
The key bottleneck for enterprise AI adoption is not better models or more context, but "better knowledge."
Practice case: background 'dreaming' agents account for 20%-50% of runs
After the article was published, Elias Lumer shared his team's internal agent practices, which Chase reposted. Lumer revealed that his team divides context into different types such as skills, system prompt, user memory, and global memory, most of which are autonomously updated by background agents.
Lumer specifically noted that the share of background/"dreaming" agents and automation continues to rise, reaching 20%-50% of total runs, with trigger intervals ranging from every 1 day or 7 days down to as short as 5 minutes.
| Context type | Update method |
|---|---|
| skills | Mainly autonomously updated by background agents |
| system prompt | Mainly autonomously updated by background agents |
| user memory | Mainly autonomously updated by background agents |
| global memory | Mainly autonomously updated by background agents |
In addition, data retention policies affect real-time knowledge pointers: data an agent reads from MCP may only be valid for 7 days, so it needs to be stored in user memory/wiki in advance as a backup.
Beyond the points overlapping with LangChain's view, Lumer's team also does something differently: they treat the curator agent (knowledge curation) and the main agent as the same agent, rather than two separate roles.
Developer's question: building tools or training replacements?
During the discussion, developer Dmitry Grinko posed a pointed question about the article: as an employee, I can't help but ask — are we building better tools for ourselves, or training our own replacements? Chase reposted the comment, showing the resonance the issue has stirred among practitioners.
Are we building better tools for ourselves, or training our own replacements?
The question reflects a deeper tension in enterprise agent adoption: while the knowledge factory boosts AI effectiveness, it may also reshape the relationship between humans and tools. The discussion is still fermenting on social media.
What to watch
Knowledge management for enterprise agents is becoming a core adoption issue. How to balance knowledge accumulation with employee value, and how to design trigger and update mechanisms for background agents, will be directions the industry continues to explore.
