Retrieval tools still have a place
Giving an agent filesystem tools works well, but it isn't the full story. In many of the domains I work in, I've had good results building custom tools for vector search or TF-IDF and letting the harness decide how to use them.
During development, it helps to have the model look at its traces: what did it do, and where did it fail to use the retrieval tools well? That gives you something specific to improve.
Match the retrieval tools and rules to the problem, and have a systematic way to evaluate whether the results are useful.
Originally posted on X. Lightly edited for reading here.