Tag
This paper introduces IntentFlux, an executable benchmark for measuring intent drift in multi-turn LLM agents, and StateForge, an explicit active-requirement maintenance method that partially improves task scores when users change their minds mid-dialogue.
A proposed workflow for AI coding agents that emphasizes brainstorming and boundary enforcement before code editing, seeking community feedback on its utility.