Definition
Plain language
A system that designs the best workflow for an AI to use on a new task in one shot, by transferring patterns from past tasks.
As stated in the literature
An amortized agentic workflow synthesizer that distills structural priors from prior search trajectories and produces a complete workflow for a new task in a single LLM call.
Why it matters: It avoids the heavy per-task search normally needed to design good agent workflows, making bespoke workflows cheap enough to use routinely.
For example, given a new data-analysis problem, SWIFT can draw on past search trajectories to emit a planner-executor-critic workflow in one shot.
Heard on the show
“The system that does this is called SWIFT.”Episode 013 — Why Search Keeps Rediscovering the Same Workflow, and What That Means