Showing an agent one tier of detail a turn too early costs up to 23 accuracy points

arXiv 2608.28027·medium signalXBlueskyLinkedInCopy link

String (arXiv 2608.28027, Aug 28) is an open runtime that moves tool knowledge out of the agent's context into a layer that renders it back one view at a time as markdown. A single String-Flavored Markdown document declares an app's views, typed actions, navigation and credentials behind two verbs, /open to see and /act to do, and the same document serves styled HTML to browsers and raw markdown to agents. The staging results are the transferable part: disclosing a tier of detail one turn early costs up to 23 accuracy points, and correct staging drops wrong-action selection from 28% to 2%. On an 87-task benchmark across six models, converting curated skills into on-demand String apps matched aggregate success (+1.3pp) using 33.5% fewer tokens, with the resident interface a constant 53 tokens regardless of catalog size. Privilege follows provenance: a remote page may call HTTP but never the shell.