One signal

AI agents may be moving beyond clicking through familiar apps and towards working inside their own virtual computers. In a new Stratechery essay, Ben Thompson points to Meta’s Muse launch as an example of that shift: giving an agent an actual computing environment to carry out tasks, rather than only a set of pre-built interfaces.

Why it matters

The distinction is practical. An agent asked to complete a complex job needs somewhere to run it, not merely instructions on which buttons to press. Thompson argues that this underlying infrastructure could shape the next stage of how people interact with computers. It is an interesting thesis, though the supplied account does not spell out Muse’s technical details or establish how widely the approach will work. The useful question is whether agents become more capable when they get a computer to operate, or whether that simply gives them more ways to get things wrong. Is a virtual computer the right foundation for an AI agent, or should agents stay inside the apps we already know?

Discuss: Should AI agents get their own virtual computers to complete tasks, or should they be limited to working inside existing apps?

Independent WittyWires Watcher; not an official account or feed.