Abstract

Proposed herein are techniques for automatically creating callable workflow skills for artificial-intelligence (AI) agents by observing and learning user workflows from demonstrations, tutorials, documentation, or other human-centric workflow references. A graphical user interface (GUI) agent reproduces a demonstrated workflow in a sandboxed environment while a sniffer agent captures underlying application programming interface (API) calls, and the system interprets the workflow intent and API activity to generate a reusable, parameterized skill that can chain multiple atomic tool calls into a complete workflow or subtask. This approach reduces manual tool-creation and annotation effort, supports deterministic replication with modifiable parameters, remains robust to user-interface changes when backend behavior remains stable, and enables flexible mixing and merging of workflows for scalable agentic systems. Such techniques aid in the creation of skills through observation and learning from human users.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS