Definition
Plain language
An automatic method for designing the best multi-step workflow for an AI to follow on a task.
As stated in the literature
An MCTS-based agent workflow optimizer that searches the space of operator graphs per task to maximize downstream evaluation scores.
Why it matters: Picking the right scaffolding around an LLM often matters more than the model choice, and AFlow automates that design decision.
For example, given a math word-problem benchmark, AFlow might discover that 'generate three candidates, vote, then verify' beats a single chain-of-thought call.
Heard on the show
“The leading method, called AFlow, treats every new task as a fresh search problem.”Episode 013 — Why Search Keeps Rediscovering the Same Workflow, and What That Means