Cursor's agent leverages semantic search to improve code retrieval and accuracy when responding to user prompts. Unlike traditional regex-based tools like grep, semantic search allows the agent to interpret natural language queries, such as “where do we handle authentication?” This capability is backed by a custom-trained embedding model and indexing pipelines designed for quick access to relevant code segments. The results are compelling: the agent achieves an average accuracy increase of 12.5% in answering questions, with variations ranging from 6.5% to 23.5% based on the model used. In evaluations using the Cursor Context Bench dataset, semantic search consistently outperformed traditional tools across various model configurations. An A/B test further highlighted the impact of semantic search on user experience. In scenarios where agents utilized semantic search, code retention improved by 0.3%, with a more significant 2.6% increase in larger codebases containing over 1,000 files. Conversely, the absence of semantic search led to a 2.2% rise in dissatisfied user requests, indicating that agents relying solely on traditional search methods required more follow-ups for corrections. The effectiveness of the semantic search stems from the custom embedding model, which uses real agent sessions as training data. By analyzing the search traces during coding tasks, the model learns what information should be prioritized. This approach contrasts with generic code similarity methods, providing a more tailored and effective retrieval process. The combination of semantic search and traditional tools like grep seems to yield the best results, especially in extensive codebases. Cursor continues to refine these tools as their models evolve.