Four cognitive operations and a flywheel: what agents should actually do to knowledge
Most agent systems answer one question: how do I get the model to write something for me? That is the wrong first question. The right one is: what kind of cognitive work should the agent do at all? After building and running a research system that has processed 7,243 documents into 138,000 knowledge nodes, 588,000 links, and 966 reports, I can tell you the framework that survived contact with…
When building an agent system for knowledge work, the focus should not be on how to get the model to generate text, but rather on the kind of cognitive work the agent should perform. The framework that has proven effective in production is comprised of two parts: four cognitive operations and a flywheel.
The four cognitive operations are: systematize, theorize, academicize, and systematize again as architecture. Systematize involves organizing scattered material into a coherent structure without creating new claims. Theorize connects specific observations to general principles, which is the most challenging operation due to the need for restraint and evidence-based connections.
Academicize brings scholarly rigor to raw arguments through citations, methods, evidence chains, and counter-arguments. The final operation is to systematize again as architecture, turning working pieces into a reproducible system where the agent writes code to maintain the system.
The flywheel describes how knowledge compounds through a series of steps: documents are turned into structured knowledge nodes, extracted from raw documents, and then used in research, followed by discovery of patterns and gaps, and finally challenged to ensure the discoveries are not just correct but also new. This process feeds back into new research, and each turn increases the knowledge base.
The gate question must be answered before any batch of work starts, ensuring alignment with the design and preventing the production of unnecessary output.
The distinction between an agent as an editor, not an author, is crucial. The agent should receive curated material nodes and be told to organize and digest the core arguments into a coherent narrative, while the agent does not generate text. Measuring trust and validity is essential, as trust is measurable through factors like citation reconciliation, method honesty, measurement-window validity, and reporting of absent events.
In conclusion, if you are building an agent system for knowledge work, prioritize classifying cognitive operations before writing prompts, treat the agent as an editor rather than an author, and incorporate gate questions before every batch of work to ensure alignment with the design. By following these principles, you can create a system that generates useful knowledge rather than impressive-looking output.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
