From intent to action.
Agents are slow because they think one step at a time. Weft is the runtime that lets a model lay out everything it needs at once — and runs it together.
The bottleneck
Thinking, one step at a time.
Today's agents loop: think, act, observe, repeat — serial by construction, burning time and tokens on the same reasoning over and over.
The idea
Emit the whole graph, once.
Weft lets the model declare everything it needs up front. A runtime runs it in parallel, reuses repeated work, and splits the load between server and browser — so the work finishes together, not in line.
Designed for
Faster, by design.
The architecture targets a step change, not a tweak. These are design goals — being built and measured, not yet benchmarks.
5× faster — target
Parallel execution instead of a serial loop.
60–80% fewer tokens
No re-reasoning the same steps over and over.
1 graph, not a loop
Declare it once; the runtime does the rest.
Observable by default
See what ran, when and why — no black box.