Google-led team introduces StreamVibe video agent

StreamVibe is designed to help during activities including cooking and assembling furniture, where an agent may need to decide whether to offer guidance.[1] Its research team includes researchers from Google, Stanford, Oxford and several U.S. universities.[1] The paper separates user intent into explicit, implicit and dynamic categories.[2] That distinction matters to the system’s stated aim: infer what a person needs even when they have not directly asked, then adjust as their intentions shift.[2] The Decoder reported on the research on Nov. 1, 2026.[1]

StreamVibe-VL makes a decision every second about whether to remain silent or respond.[1] A second component, StreamVibe-Harness, maintains user intentions as persistent states using hierarchical memory.[2] The team’s StreamVibe-Bench contains 4,100 evaluation samples and 85,000 training samples, based on 183.4 hours of video.[2][3] Its evaluation set includes 600 explicit-intent samples, 2,000 implicit-intent samples and 1,500 dynamic-intent samples.[3] Four reviewers had to approve a sample unanimously for it to be retained.[3]

The researchers report that StreamVibe-VL improved response timing and answer quality by 10.7% over existing baselines, while StreamVibe-Harness improved proactive interaction performance by 12.3%.[2] The paper also says StreamVibe-VL preserves strong general video-understanding capabilities.[2] These are reported research results, not evidence that the system is ready for every work setting. In The Decoder’s account, Google’s proprietary Gemini 3.8 Flash outscored StreamVibe-VL in several tests.[1] The comparison underscores that the reported gains do not mean StreamVibe led every evaluation.[2][1]

The Decoder reports that performance across all tested systems plummeted in continuous video evaluation as errors compounded.[1] A key content metric for responses fell by as much as 61.5% when systems relied on their own prior history.[1] Models rarely chose the right moment to respond in videos longer than five minutes, and the researchers say they struggle particularly with instructions that change during long videos.[1] Examples included JoyAI saying a GIMP image had been saved before the user opened it, and StreamVibe-VL advising someone to cut off a jam-roll’s bread crust first.[1] The researchers caution that real-world situations may be more complex than their test setup.[1]