Cloudflare’s Post

The architectural bifurcation between fast, deterministic 'System One' extraction models and compute-intensive 'System Two' reasoning engines is the key to making production agent pipelines economically and latency viable.

The interesting part is collapsing extraction, computation, and interpretation into one layer. If the model can reliably turn messy inputs into structured answers, a lot of multi-step pipelines get much simpler.

Possibilities! Now combine that with Clef - Clairic extracts uncertain structured data; Clef decides based on policies. Amazing!

Extraction from unstructured inputs is where most enterprise AI value is hiding: contracts, invoices, tickets, claims. A System One / System Two split for models matches how ops teams already route work, fast path for the routine, deep path for the ambiguous.