Andrej Karpathy's vision for Tesla's autopilot to improve autonomously has inspired ideas for chat-based software like Meridian. The development approach at Meridian focuses on gathering user insights to enhance the app, transforming feedback into feature requests or bug reports that engineers address. This iterative process typically takes from an afternoon to several weeks, depending on the complexity of the request. However, the introduction of coding agents is speeding up this loop, enabling the team to turn well-defined requests into pull requests within hours or days. For instance, users expressed a desire for scheduled trading actions in the app, prompting the product team to draft specifications and the engineering team to implement a price monitoring service. Coding agents can now handle detailed specs and rapidly experiment with different implementations, streamlining the experimental loop. The efficiency of this system hinges on equipping agents with the same tools as human engineers, allowing them to inspect logs and test features in development environments. The article outlines potential methods for automating the feature pipeline, emphasizing the need for systems that can autonomously generate bug reports and understand user demands. User requests and telemetry data from legacy systems can provide valuable insights. As these automated systems become more capable, they may identify user wants, such as scheduled trading actions, based on chat interactions. While humans will still play a role, their tasks will shift toward defining broader objectives and guiding agents in decision-making processes. The ultimate goal is a closed-loop software development system that continually adapts and improves based on user feedback, making the development process more efficient and aligned with user needs.