A-TIER repo
Prime Agent — self-rewriting harness lifts Opus 5 to 95.5% on ARC-AGI-3
100k ft — What is this?
Prime Agent — self-rewriting harness lifts Opus 5 to 95.5% on ARC-AGI-3
50k ft — Why care?
Prime Intellect released Prime Agent, an open-source coding harness built on a Recursive Language Model (RLM) framework that rewrites its own prompts, memory, tasks and scaffolding at runtime based on failures. Opus 5 alone scores 30% on ARC-AGI-3; with the harness, 95.5%, a hair over the 95.4% human baseline. The real story is test-time compute and self-modifying scaffolding.
10k ft — Key takeaways
• A 0.1% margin over human baseline is noise, not superhuman.
• Harness design can triple benchmark scores without a new model.
• Let scaffolding adapt to observed failures at runtime.
5k ft — Tools & techniques
Prime Agent
AI agents, code generation, recursive language models, self-modifying scaffolding, test-time compute scaling
1k ft — Deep-dive / try it
🟢 Comment “AGENT” — I’ll send the repo
The Prime Agent headline is hype. Here’s what actually happened.
Prime Intellect released Prime Agent, an open-source coding harness that uses a self-improving Recursive Language Model (RLM) framework. Headline claim: 95.5% on ARC-AGI-3, above the human baseline of 95.4%.
Reality:
- Opus 5 alone: 30% on ARC-AGI-3
- Opus 5 + Prime Agent harness: 95.5%
- Margin above human: 0.1% (statistical noise)
- Real story: test-time compute scaling + self-modifying scaffolding
The breakthrough isn’t the model — it’s agent infrastructure that lets the harness rewrite itself at runtime based on what’s failing. That’s worth paying attention to.
🟢 Comment “AGENT” and I’ll send you the repo
Transcript excerpt
One harness took Opus 5 from 30% to 95% on the ARK AGI 3 scale, so we just got the best self-improving harness out there. Extremely token efficient and it makes your AIs much better. Which tells you the differences making your AIs more efficient is the way it scaffolds. The harness that operates your AI makes it a lot better. The difference is the model didn't just get better, but the way it scaffolds. This harness does something completely different. While it's running it could rewrite its own prompts, memory, and its tasks, and modifies its own harness for self-improvement. This is one of the most revolutionary things that we have currently, so remember, it's not the way your LLMs improve, but it's also the way your harness operates them.