Kimi CLI Former Author: The more powerful the model, the thicker the Harness

According to Ooda Beating monitoring, Raft founder and former author of Kimi CLI, Richard Qian, put forward a counterintuitive proposition: the stronger the model, the thinner the Agent harness may not necessarily be, but instead could become thicker.

Things like System Prompt, dedicated Tools, and Subagents may gradually disappear. As models become smarter, many fixed processes can be directly handed over to the model, eventually leaving only a Bash tool.

However, once Agents start working long-term and collaborating, new issues immediately arise: who is responsible for what, how far along tasks are, how to synchronize context, and how multiple Agents can avoid duplicated work and conflicts. All of these need to be resolved by systems outside the model.

After leaving the dark side of the moon, Richard, in his work on Raft, is actually addressing this layer. It brings together Agent such as Claude Code, Codex, Kimi CLI, into the same workspace, allowing them to chat in channels, claim tasks, read message history, and then hand off work to each other.

The model is eating away at the old harness, but as Agents can do more, the external collaborative system will also become more complex.