Prime Agent Testing - Why is it said that in the RSI era, a 0.1% model gap may become an insurmountable gap? The Prime Agent released by Prime Intellect is a runtime and testing framework that enables models to work long-term, call tools, organize sub agents, and improve their own processes. The test results are quite astonishing: Claude Opus 5 scored approximately 30.2% on ARC-AGI-3 in native Harness; After switching to Prime Agent, the score increased to 95.5%, slightly higher than the baseline of 95.4% for human experts. GPT-5.6 Sol has further increased from 13.3% in official Harness and 38.3% in optimized context to 78.3%. Prime Agent has validated a straightforward truth through testing: once the model capability gap enters a long task, it no longer grows linearly, but accumulates exponentially along the execution chain. Prime Agent turns context into programmable data through RLM. The historical records, task status, and intermediate results are stored in a persistent Python environment, and the model autonomously decides what code to write, what information to extract, and when to call the tool. Strong models are easier to establish correct state representations, while weak models may preserve erroneous assumptions from the beginning. And each subsequent step is built upon the cognitive structure previously formed. It also allows models to build persistent multi-agent teams. One agent studies rules, one writes code, and the other checks for vulnerabilities. The Prime Agent is responsible for communication, backend operation, and state preservation, while the model decides how to divide labor, when to parallelize, and how to trust. This is equivalent to incorporating the organizational capability of the model into Scaling: The model that can divide labor gains parallel benefits; A model without division of labor will only allow errors and token consumption to expand together. Continuous Harness allows the model to review the execution trajectory, write effective methods into Memory and Skill, modify Prompt or adjust sub agent configurations. If the judgment is correct, successful experiences will continue to be reused; Attribution errors and occasional biases may also be solidified. The above three aspects all rely on the model's continuous error correction and optimization, as well as its multi-step execution ability. This is why a 0.1% gap may become a gap. Assuming that the reliability difference between each step of the two models is only 0.1%, a single comparison is almost imperceptible; The gap will continue to compound after hundreds of steps of task execution, combined with memory management, agent coordination, experience extraction, and process modification. A stronger model not only performs slightly better in this round, but also leaves better tools, memories, and organizational structures for the next round. Create a positive compound interest effect. The significance of the Prime Agent framework and testing lies not only in using a long-range task approach to further test model capabilities, but more importantly, in demonstrating the huge differences that a weakly leading model may produce after multiple iterations in the era of recursive self iteration (RSI) of the model. Prime Agent also tells us that there are still many new potentials to be explored in scaling law at the agentic level.