Beyond Tool Selection: Why LLM Agents Still Struggle With How to Use Tools

Developers building LLM-based agents face a recurring challenge: deciding which tools to expose and when, without overwhelming the model's context window. Early approaches like static tool sets and search-based discovery proved inefficient, either flooding the context or causing the model to waste time guessing search terms. Dynamic tool recommendation improved task completion but disrupted cache reuse and raised inference costs significantly. A more effective pattern emerged — selecting capability groups at the start of a work unit and keeping them stable throughout, rather than swapping tools every round. However, having the right tools available did not prevent models from using them inefficiently, such as reading entire files when only a few lines were needed or skipping result verification.