This is a problem I’ve noticed with long-running agents too. More context helps, but without a clear budget, the agent carries more noise around.