>It just slurps in all of the code, documentation, and supporting material, and knows everything. It's an immediate expert.

In practice this is not how agentic coding works right now. Especially for established projects the context can make a big difference in the performance of the agent. By doing simpler tasks it can build a memory of what works well, what doesn't, or other things related to effectively contributing to the project. I suggest you try out OpenClaw and you will see that it does in fact learn from practice. It may make some mistakes, but as you correct it the bot will save such information in its memory and reference that in the future to avoid making the same mistake again.