According to monitoring by Dongcha Beating, Raft founder and former Kimi CLI author Richard Qian has proposed a counterintuitive view: the stronger the model, the thinner the agent harness may not necessarily be—in fact, it may become thicker. System prompts, specialized tools, and subagents may all gradually disappear. As models become smarter, many fixed workflows can be handed directly to the model, eventually leaving just a single Bash tool. But once agents start working over the long term and collaborating with each other, new problems immediately surface: who is responsible for what, how far along a task is, how to synchronize context, and how multiple agents can avoid duplicating work and conflicting with one another. These must be solved by systems outside the model. After Richard left Moonshot AI, he built Raft precisely to fill this layer. It brings agents such as Claude Code, Codex, and Kimi CLI into the same workspace, allowing them to chat in channels, claim tasks, read message history, and hand off work to one another. The model is eating the old harness, but the more agents can do, the more complex the external collaboration system will become.