Prime Intellect released Prime Agent, a free and open-source command-line coding agent that can rewrite its own instructions, saved notes and helper agents based on what it learned during a run. Paired with Anthropic's Opus 5, it scored 95.5% on the ARC-AGI-3 reasoning benchmark, slightly above the human expert score reported by the benchmark's organisers. No model has been trained specifically for this harness, and the reported results come from the team's own testing.
What changed
Coding agent harnesses used fixed tool-calling schemas and hand-written prompts, skills and sub-agents set at design time.
What it unlocks
Running an open-source coding agent that edits its own prompts, skills, memory and sub-agents from its own run history, and that can work unattended under an explicit token and turn budget.
- 95.5% on ARC-AGI-3, best@1
- human expert baseline 95.4%
- 183/183 levels complete
What you need to act on it
- command-line install
- access to a frontier model, open-weights or commercial
- engineering setup to run agents locally
Sources