Oossa

Vanguard checks AI agent drafts against screenshots

Researchers introduce a training-free system that lets a smaller model propose actions and uses screenshots to decide when those drafts are safe to keep.

NoteBy Published by Oossa: 1 min read

Researchers have introduced Vanguard, a system for checking action suggestions from smaller AI models before a web, phone or home agent uses them. It looks at the proposed step and the screenshot, then either keeps the draft or asks a larger model to write a replacement.

In the researchers’ tests, a 2-billion-parameter draft model judged by Vanguard was at least as accurate as its 8-billion-parameter target, and more accurate on two named benchmarks. The larger model wrote only a minority of steps.

Why it matters

For teams building website or phone agents, Vanguard could mean fewer calls to a larger model without sacrificing accuracy in the tested tasks; whether that holds in deployed products remains unknown.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.