Computer-use agents are interesting because they can work through interfaces that already exist, rather than wait for every organization to provide a new integration. On Practical AI, Demetrios Brinkmann describes multi-account briefings, flight preparation, inbox-triggered check-in, a government appointment, and a LinkedIn workflow that discovers an official data export and resumes after it arrives. These are personal reports, not measured reliability results, and the limits are revealing: he still approves the final payment, worries about repeated small purchases, and calls unrestricted enterprise computer use a security nightmare. The same browser session that removes integration friction can expose an enormous amount of authority. He also raises a model-harness compatibility problem—the harness is the software supplying tools and control flow around the model, while a model may have learned assumptions about a different environment. The closing commerce discussion is sharper than the usual agent-swarm forecast: if browser automation already completes the task, what forces a more complicated agent-to-agent internet into existence? And if an assistant becomes a marketplace or payment intermediary, can it remain a trustworthy buyer’s representative while profiting from steering purchases? The useful dividing line is not chat versus websites, but bounded delegation versus an opaque intermediary with conflicting incentives.
“you don’t have to wait for the other person to set up an MCP server” — Demetrios Brinkmann, 24:46
“does any of this even matter? Is there going to be a forcing function to create this reality if with browser use, I can already get what I want?” — Demetrios Brinkmann, 48:34
Based on the full publisher transcript; quotations checked against its timestamped captions. The proposed shopping architectures are thought experiments, not announced products.