What they did

The authors built String around two agent-facing operations: /open renders the current view, while /act validates and executes an action. The runtime handles tool discovery, state and credentials, and can expose apps, files, shells, websites and legacy HTML through a common Markdown interface; the same source can render styled HTML for people.

They evaluated the system on an 87-task benchmark in which each task was paired with curated procedural skills. Across six models, from frontier systems to smaller models, they compared those skills with procedures operationalized as on-demand String apps. They also varied when interface details were disclosed to test the causal effect of staged information.

Key findings

• String produced comparable aggregate task success to curated skills, improving it by 1.3 percentage points overall.
• Among completed episodes, String reduced token use by 33.5%.
• The permanently loaded interface occupied 53 tokens regardless of the number of available applications.
• Revealing one tier of detail a turn too early reduced accuracy by as much as 23 points; appropriate staging lowered wrong-action selection from 28% to 2%.

Why it matters

The results support treating agent interfaces as a context-management layer rather than placing every tool definition in the model prompt. A compact, progressively disclosed grammar could make large tool catalogs cheaper to navigate while centralizing action validation, state handling and restrictions on secrets or shell access.

Caveats

The evaluation covers 87 tasks paired with curated skills, so it does not establish performance across arbitrary applications or fully autonomous, long-running workflows. The abstract reports three months of production use but provides no scale or controlled comparison for that deployment, and the stated privilege rules do not by themselves demonstrate resistance to all prompt-injection or credential attacks.