The research centers around developing a self-driving codebase, primarily using a web browser project as a benchmark. The initial goal was to create a browser engine capable of rendering web pages without JavaScript. Early attempts with a single agent struggled due to the complexity of the task, leading to a fragmented approach where agents couldn't effectively communicate or collaborate. Upgrading to GPT-5.1 and GPT-5.2 models improved instruction-following, but a single agent still struggled with the workload. To enhance productivity, the team shifted to a multi-agent system, initially using a shared state file for coordination. This approach failed due to issues with locking and contention, which slowed down the process significantly. After analyzing the problems, they restructured the system by assigning distinct roles: a planner to outline tasks, an executor to oversee implementations, and workers to carry out specified jobs. This division allowed for better accountability and streamlined operations. The design evolved further; the executor assumed both planning and execution roles, leading to a more dynamic system that could adapt to changing conditions. Freshness mechanisms, including regular updates to notes and self-reflection prompts, were implemented to prevent drift in agent performance. Despite progress, the executor sometimes exhibited erratic behaviors due to juggling too many responsibilities at once. The final design aimed to simplify roles while maintaining the flexibility and responsiveness of the system, incorporating lessons learned throughout the research process.