I pretty much never clear my context window unless I'm switching to entirely different work, seems to work fine with copilot summarizing the convo every once in a while. I'm probably at 95% code written by an llm.
I actually think it works better that way, the agent doesn't have to spend as much time rereading code it had previously just read. I do have several "agents" like you mention, but I just use them one by one in the same chat so they share context. They all write to markdown in case I do want to start fresh if things do go the wrong direction, but that doesn't happen very often.