Prime Agent completes 178 of 183 ARC levels

Prime Intellect released Prime Agent, an open-source coding harness whose published ARC-AGI-3 run completed 178 of 183 levels across 24 of 25 environments for a 95.24% score. Prime Intellect’s launch post describes a persistent Python environment for programmatic tool and subagent calls plus a continual harness that can revise supplemental prompts, memory, skills, and subagent specifications; the ARC Prize scorecard independently records the run, although one environment remained incomplete and the launch benchmark is not a coding-production evaluation. The result reinforces the evidence from same-model harness comparisons: builders evaluating agents should version and test orchestration, context management, and stopping rules alongside the model rather than treating the wrapper as neutral plumbing.