Harness Will Not Disappear as Models Improve, Says Ex-Kimi CLI Lead

According to a report by Leiphone.com, Qian's comment came amid opposing trends: minimal-agent products like Pi use fewer than 1,000 tokens and just four basic tools, while other developers are building increasingly complex agent swarms. Many believe Pi's success proves that models alone are enough. Qian disagrees, saying those who predict harnesses will be trained into the model “have never done harness work.” He offers three reasons: harnesses pave the way for models to learn collaboration; higher intelligence demands more complex tools, just as humans invented laws and the internet; and single tasks may be handled by models, but complex business logic involving long processes and multiple agents cannot rely on a model's memory alone.

He also addresses the misconception that stronger models make harnesses thinner. While prompt patches and rigid retry mechanisms may fade, new harness needs emerge at a higher level—multi-agent communication, cross-session state synchronization, proactive memory management, dynamic permissions, and integration standards. Qian describes the relationship as yin and yang: some local harness functions become internalized by models, while new upper-level requirements grow outward.

Qian's view comes from his experience building Kimi CLI from scratch. At one point, his team removed the dedicated subagent scheduling code and native parallel control, letting the model write scripts to handle those tasks. The experiment worked, but only once model capability reached a threshold. Then the team had to face new problems: how do multiple agents communicate, hand off tasks, and agree on protocols? That exploration later became the technical foundation for Raft, his current startup.

At Raft, Qian has deliberately discarded the single-agent runtime. Instead of writing agent loops or encapsulating basic tools, Raft treats existing products like Claude Code and DeepSeek as team members. Its focus is on the upper-layer harness: giving each agent its own process, memory, and habits; organizing work through channels with task claiming, handoff, and audit trails; building a cross-vendor communication protocol so different models talk in one workspace; and creating a shared workspace where humans and AI work together. Pricing is notable: each agent counts as 0.1 of a human seat.

Qian's final observation: “When everything is a harness, it becomes invisible.” Just as people do not constantly mention company rules as the harness of society, a mature AI harness will blend into daily workflows like electricity or water. For practitioners, the real question is not whether models or harnesses will win, but how to build durable advantages in higher-level collaboration and complex business scenarios.