ANIS synthesizes agent-native desktop interfaces: +3.8pt success, 73% fewer tokens
tencent · hf · 2026-10-14
Desktop UIs are built for humans, forcing agents to parse visual controls or compose low-level APIs. The study builds DeskIF-Bench (255 artifact-verified tasks, six apps) and shows newer releases consistently outperform older ones, identifying discoverability, granularity and observability as key interface properties. The ANIS framework synthesizes agent-native interfaces from installed app capabilities: on 368 tasks across eight apps, atomic interfaces lift success by 3.8/3.5 points over native SDK use for DeepSeek V4 Flash and Claude Haiku 4.5 while cutting online token usage 73%; goal-level commands push success to 74.5% and 67.4%.
More from coding & agent
- Playable Multiplayer Game Embedded in a Tweet Sparks Red vs Blue Pill Battle — tibo_maker · 2026-10-14
- One AI audit of curl produced 24 CVEs; 18 fixed in 8.23.0 today — aidangomez · 2026-10-14
- Asked Fable to review code, it spawned 6 subagents and burned through the 5-hour limit — Aryvyo · 2026-10-14
- Using Obsidian as your agent's knowledge base: distill every conversation into md notes — mazzaTalk · 2026-10-14
- Two small LLMs on a Galaxy S21 complete an Amazon shopping task locally in 51 minutes — Mean-Standard7390 · 2026-10-14
- Tech in Five: a quick primer on what Computer Use actually is — JohnSavill · 2026-10-14