This resonates heavily with our experience. We ended up using one agent + actively managed context, with the smartness baked into how we manage that context for that one agent, rather than attempting to manage expectations/context across a team of agents.