Haven’t read the thread, and I’m still sleepy, just woke up. But the premise that the “agent has to learn everything in the project from scratch” is false. It doesn’t need to understand the entire project to implement a new feature or fix a bug. This is obvious true, because much can be inferred from a single file from the interfaces it interacts with. And the more self-contained your files and modules are, the less context it needs. Of course, some tasks require a broader understanding of the codebase. But that can be rediscovered as needed. This is something I realized pretty early on. And it’s still one of the things I actively control as the human in the loop: I decide what context the agents get, and what they need to know. A well specified prompt will make the agent state whether more context is needed, but the agent can figure that out by itself anyways. I’d go further and argue that “knowing everything about the project” is detrimental. The underlying architecture is still the same, and everything in the context competes for attention. Things unrelated to the task injected into the context can only be detrimental.