The solution is simple, and it's what's already done with search by proprietary LLMs: reasoning happens on the LLM vendor's servers, tool use happens client-side. Whether for search or "computer use", the websites will register activity coming from the user's machine, as it should be, because LLMs act as User Agents here.
Of course, already with LLM-powered search we see growing number of people doing the selfish/idiotic thing and blocking or poisoning user-initiated LLM interactions[0]; hopefully LLM tools following the practice above will spread quickly enough to beat this idea out of peoples' heads.
--
[0] - As opposed to LLM company crawlers that scrape the web for training data - blocking those is fine and follows the cultural best practices on the web, which have been holding for decades now. But guess what, LLM crawlers tend to obey robots.txt. The "bots" that don't are usually the ones performing specific query on behalf of users; such bots act as User Agents, neither have nor ever had any obligation to obey robots.txt.
Of course, already with LLM-powered search we see growing number of people doing the selfish/idiotic thing and blocking or poisoning user-initiated LLM interactions[0]; hopefully LLM tools following the practice above will spread quickly enough to beat this idea out of peoples' heads.
--
[0] - As opposed to LLM company crawlers that scrape the web for training data - blocking those is fine and follows the cultural best practices on the web, which have been holding for decades now. But guess what, LLM crawlers tend to obey robots.txt. The "bots" that don't are usually the ones performing specific query on behalf of users; such bots act as User Agents, neither have nor ever had any obligation to obey robots.txt.