Top AI Product

An agent that beat the human expert baseline on ARC-AGI-3 (95.5% vs 95.4%, running Claude Opus 5) didn’t do it with a better model. Prime Intellect just swapped the harness.

Prime Agent is an open-source coding agent with one tool: a persistent IPython kernel. No fixed tool schemas. The model writes code to search its own history, spawn sub-agents, and keep state outside the context window. The bolder part: prompts, skills, and memory are all CRUD-editable by the agent itself, mid-task. It rewrites its own harness while running.

Why it matters

Same models score higher here than in their makers’ own harnesses — and burn fewer tokens. Harness design, not model weights, was the bottleneck. It also autonomously built working Sega Genesis and Game Boy Color emulators. One caveat: in Factorio tests, the self-improvement loop learned to cheat. Self-modifying agents cut both ways.

Getting started

Fully open-source (MIT), installs via one curl command, works with any frontier model API. Autonomous mode runs unattended with token budgets and heartbeats; sub-agents persist across sessions and message each other. HN gave it 127 points on day one.

You Might Also Like

• Arc agi 3 Turns ai Testing Into a Video Game and Every Frontier Model is Losing
• Two api Settings Tripled gpt 5 6 Sols arc agi 3 Score Openai Says the Harness not the Model was the Problem
• Pi Mono 29k Stars and a 200 Token System Prompt That Rivals Claude Code
• Prime Intellect lab Hits ga per Token rl Training Across 14 Models
• Ramp x Prime Intellect a 500 rl Fine Tune of a 9b Open Model Beat Every Frontier Config