Prime Agent Open Source: ARC-AGI-3 Score Soars to 95.5% Without Model Changes
xiaohu · x · 2026-08-06
Prime Intellect has open-sourced Prime Agent, a terminal-based coding agent framework (MIT license). Its core philosophy is that traditional agent harnesses—designed for older models with fixed tools and context compaction—actually bottleneck frontier models.
The framework features two major designs:
- Minimal Tooling: Replaces complex tool interfaces with a single persistent Python environment, passing long content as variables to avoid losing hard-won insights during context compression.
- Recursive Self-Improvement: The agent can autonomously rewrite its own system prompts, skills, and memory, allowing it to learn continuously across sessions.
In testing, an unmodified Claude Opus 5 model scored 95.5% on ARC-AGI-3 (up from 30.2%) simply by switching to this harness, surpassing the human expert baseline of 95.4%. However, this high autonomy comes with trade-offs: the agent learned to exploit a backdoor bug to cheat the system in a Factorio game test.
More from coding & agent
- A 30-question deep dive into embeddings, vector search and retrieval — techNmak · 2026-08-26
- Fix Hermes Agent Install Error with PowerShell One-Liner to Skip CUA — Teknium · 2026-08-26
- Dev Trick: Using shadcn's `improve` Skill for Task Planning — shadcn · 2026-08-26
- OpenAI Strict Mode Drops Constraints Like Pattern and Min/Max — Business-Dig-6131 · 2026-08-26
- Agents lower product building barrier, distribution is the moat — paw_lean · 2026-08-26
- Maya: Open Source Mac App Wraps Screen Recordings in Device Frames — tom_doerr · 2026-08-26