having worked on background coding agents at
@ctodotnew since before Claude Code was GA, it feels like I've tried out all the major strategies and techniques to manage long contexts
I'm still pleasantly surprised to find new ones appearing, like in Kimi's CLI tool, which lets the agent send a message "back in time"
my medium-spicy take is that background agents are better than interactive coding agents for a large class of coding tasks - one small but important reason is that they let you think about long context management as a system design problem rather than a strictly algorithmic one