It's more accurate to say the agents like simple, interpretable tools. Clear input / obvious output. That might be more an AST thing, keyword search, search with very straightforward filters, etc.

I don't thing they do well with search that is built for human engagement, which is a more complex tool to reason about.

▲

esafak 7 hours ago | parent [-]

It makes no different to the agent. The response from grep is a list of matches, and so it is from something more intelligent. A list is a list.

> I don't thing they do well with search that is built for human engagement, which is a more complex tool to reason about

I agree! Structured outputs are best.

	▲	shadowgovt 3 hours ago \| parent [-]
		Under the hood, LLMs are vector analysis engines. They benefit from spaces where the result of change is smooth. Adding levels of indirection and secondary reasoning to the search interface makes the results less smooth. This is one of the things humans complain about often when using e.g. Google: "I'm interesting in searching for X, but all these results are Y." Yes, because X and Y are synonyms or close topics and Google is mixing in a popularity signal to deduce that, for example, your search for `tailored swift database` is probably about a corpus of Taylor Swift song lyrics and not about companies that build bespoke Swift APIs on top of data stored in Postgres. If you're driving the process with an LLM, it's more of a problem for the LLM if it's searching the space and the search engine under it keeps tripping over the difference between "swift means a programming language" and "swift means a successful musician" as it explores the result space. A dumber API that doesn't try to guess and just returns both datasets blended together fits the space-search problem better.