The fastest model call is none

Last reviewed: July 2026.

"Filter to last 30 days." "Sort by date descending." "Expand the second panel." A naive agent system sends each of these to a language model, pays a round-trip of latency and tokens, and gets back an intent that was never in doubt. Multiply by every mechanical operation in a working session and the model becomes a tax on interactions that a pattern table could have resolved.

Fuaran's position is that the deterministic tier comes first, and the language carries the declarations that make it possible:

• Fast-path patterns live on the node. A node may declare natural-language patterns – exact phrases, slotted templates ("filter by {field}"), regular expressions – mapped to its own affordances. A resolver consults the current tree's declarations before any model is invoked; a match dispatches the action directly, in microseconds. The patterns are colocated with the affordance they trigger, so the vocabulary travels with the component rather than living in a per-application lexicon – and because a pattern's slots correspond to the action's typed payload, a successful match is structurally valid by construction.
• Known shapes resolve by structural search. A request for "a KPI dashboard with a filter" doesn't need generation if a certified pattern already exists: a signature search over a bank of known-good trees returns it deterministically – no key, no network, zero marginal cost. You can try this in the playground's pattern bank: the lookup is real and instant.
• The model handles the residue – the genuinely novel emission, the ambiguous request – with the full typed catalogue in hand. And a low-confidence pattern match falls through to the model rather than guessing: determinism where it's safe, intelligence where it's needed.

The economics matter more than the elegance. An interactive operation has a latency budget of hundreds of milliseconds; a model round-trip spends it all. A visitor's (or a product's) token budget is finite; every mechanically-resolvable utterance that reaches the model is waste. And determinism compounds with trust: a pattern dispatch does the same thing every time, which is exactly the property you want for the operations users perform hundreds of times.

There is a design lesson buried here that generalises beyond UI: "add AI" does not mean "route everything through a model". It means putting the model behind a deterministic front – typed patterns, structural search, closed value spaces (Declared, not scraped) – so its calls are reserved for the questions only a model can answer. The systems that feel fast will be the ones that know when not to think.