Here's what I think is useful in MCP today for most people: search & apply. Like the old school google one-shot ("I'm feeling lucky" button) but with a cognitive part in the middle to determine the top result by actually reading the link.

So you let <chat agent> query the search engine, then let <search ranking agent> rank the results properly (zhe algo is awfully good at determining relevance, to my surprise and dismay) and then let <chat agent but without memory> answer the original question based on the provided results.

The difference between what current search engines do in their LLM suggestion top box and this is that they don't read the linked content, but the indexed meta tags. Which is why i really like my news summarizer because it gives meaningful summaries rather than SEO-gameable results.