The strict parent of closed-source RSI has arrived: 18 agents conduct autonomous research, and Kimi K3 approaches Opus 5 with Harness

Open-source AI infrastructure company Prime Intellect proposed in a recent study:

With a multi-agent Harness, monitoring and concrete implementation can be handed off to smaller, cheaper open-source models, making "AI developing AI" cheaper and even more effective.

In the experiment, they put 18 frontier models, including Fable 5, Opus 5, GPT-5.6 Sol, and Kimi K3, into a nanoGPT optimizer speedrun.

As a result, the open-weight model Kimi K3, paired with the Prime Agent Harness, achieved 2,930 steps in the original results page.

This not only surpassed GPT-5.6 Sol's 3,042 steps, but also came within just 10 steps of Claude's flagship model Opus 5 at 2,920 steps.