Software-operating agents struggle with graphical user interfaces because screenshots miss hidden state and clicks are brittle. ASIL (Agent-Software Interaction Layer) replaces this with structured JSON observations and code-executable semantic actions, realized through file, scripting, or API access paths.
Tested across 15 applications on 300 single-application and 80 multi-application tasks, ASIL reaches above 80% success with closed models using fewer than five actions per task. In contrast, screenshot-and-click methods achieve only 6.6% and 26.6% strict success under a 50-step budget. The structured interface also improves training, with small-scale SFT raising a 2B model from 58.0 to 72.1 and a 9B model from 66.6 to 80.4, further boosted by on-policy RL to 74.4 and 82.2.