What they did
Igor Kamenetskiy, who works on SAP Analytics Cloud (SAC) projects at Deloitte, wanted to know if a coding agent could do SAC development work that the public APIs cannot, such as creating a new model with its dimensions, measures and properties. He gave an AI coding agent browser control through Playwright MCP and wrote a SAC skill with two layers: where each function lives in the SAC menus, and step-by-step procedures for tasks like building a model. The agent read a JSON file describing the model and then clicked through the SAC interface to create it. To write the skill faster he recorded himself doing each operation with Playwright's recording tools and had the agent turn the recording into Markdown instructions, which he then corrected. He does not name which coding agent he used.
How it went
The agent did create SAC models through the UI from the spec, and the recorded-steps method gave him a set of instructions for creating, changing and deleting models. Even with the skill, it was so slow that his demo video is sped up five times. Token use was the bigger problem: about four hours of work with the prototype used up his whole monthly token allowance. His conclusion is that it works but is not practical for everyday SAC development yet.
Worth knowing
Driving a GUI means the agent re-reads the page state at every step, so token cost stays high even when it already knows the route. Recording the steps yourself and having the agent write them up is the cheap way to build such a skill.