Architecture

Upstream/Downstream

The separation of research agents (describe reality) from recommendation agents (interpret and activate), preventing bias contamination in pipelines.

Upstream/downstream is the most important architectural principle in Knowledge OS agent design. Upstream agents collect evidence, encode structure, and preserve optionality. They describe what they find without editorializing. Downstream agents interpret upstream output, score it, and generate recommendations. The two contamination vectors are role blending (one prompt does both research and recommendation) and tone bias (loaded language like "concerning trend" in supposedly neutral research). The gold-standard upstream prompt is the company evidence research template. This principle cascaded to prompt-creator, quality evaluation, and the B2B prebuild template.

Where it shows up:

Agent pipeline designResearch workflowsPrompt architectureQuality evaluation

Build your Knowledge OS

90 minutes from zero to your first skill chain.