Cursor recently launched Cursor 2.0, introducing Composer, a coding agent designed to speed up code generation. According to a survey by Sonar, 96% of developers don't completely trust AI-generated code, with 61% noting that while AI can produce code that looks correct, it often isn't reliable. Despite these concerns, Cursor claims Composer can execute coding tasks four times faster than similar models, completing most requests in under 30 seconds. Building a reliable coding agent involves complex systems engineering. Cursor's coding agent architecture includes several components: a router that selects the best model for a task, an agentic coding model like Composer, and a tool harness with over ten integrated tools for various coding operations. The orchestrator manages the agent's execution loop, allowing it to reason about tasks and call appropriate tools. Context retrieval ensures the agent works with relevant code snippets without exceeding prompt limits. Cursor faced significant challenges in production, particularly with reliable editing, latency, and sandboxing. The "Diff Problem" arises when models struggle to edit existing code correctly, affecting the quality of outputs. Compounded latency issues can slow down iterations, while sandboxing is necessary to execute commands safely without risking the host system. These hurdles highlight the complexities of integrating AI into software development workflows, emphasizing the need for robust engineering solutions.