Agents should do the work

Chat that summarizes your inbox is a feature. Cross-application execution with context is a company.April 14, 2026

A lot of "AI agents" today are chat with better manners. They draft. They summarize. They suggest. Then a human still has to click across five apps to finish the job. That is autocomplete with a product page. The useful next step is cross-application execution: systems that can see the screen, understand the workflow, and carry intention through tools the way an operator would. Not a wrapper on one API. A loop that can actually finish the work. I built toward this with Clippy, a macOS agent meant to learn workflows from behavior and execute across applications. The thesis was blunt: if the computer can see what you see and act where you act, the chat box stops being the product surface. Bhupi AI taught me an earlier version of the same lesson from the opposite angle. Train a model on 36 hours of one professor, and people don't just want answers — they want the manner of the person. Contextual intelligence beats generic cleverness. Consumer scale (thousands of weekly actives, rooms at Google and Microsoft) proved the craving. Enterprise is where it becomes infrastructure. Distribution and reach outlast any single model choice. Models churn. The company that owns the execution layer and the trust graph compounds. So when I look at AI decks now, I ask one question: does this thing do work across the messy real software people already use, or does it narrate work back to them? If it only narrates, it is a feature. If it finishes work across tools, it can become infrastructure for how labor gets reallocated. I am biased, obviously. Bias built on shipping still beats neutrality that never ships.