Intel / article

Local Edits, Global Ripples: Replay-Informed Policy Adaptation for Workflow Synthesis

Source: arxiv.org
Author: Manqing Mao, Hong Wang, Samson Koelle, Jie Yuan, Zhuoer Wang, James Feng, Yanjun Lin, Daniel Edmiston, Nikki Lijing Kuang, Zhecheng Sheng, Wei Niu
Date: Published Sep 14, 2026

Context

A prompt change that helps an alone can become harmful when combined with earlier changes. Researchers at Amazon built RIPPLE to test that interaction while improving agents that generate executable workflows. It identifies failures, proposes corrections to specific instruction sections, then reruns candidate changes on top of the already accepted edits. The paper keeps the underlying model and tools fixed.

One experiment shows why retesting matters: a change that limited resource lookups initially helped, but later allowed stale identifiers to survive when combined with another instruction. RIPPLE rejected it after testing the combination. The study's synthetic workflow supports evaluating the assembled instruction set whenever an agent permanently updates its rules.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source arxiv.org