Prime Intellect open-sources Prime Agent, a self-improving harness

MIT-licensed, it runs sub-agents as async function calls in one IPython kernel, rewrites its own harness via /refine, and edged the human expert on ARC-AGI-3.

Nowline AUG 8 12:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • One curl, and you have a self-improving agent

    Prime Intellect's Prime Agent installs on Linux or macOS with a single curl command and runs immediately — no container required. It's fully open under MIT, so you fork the harness itself, not just call an API.

  • Sub-agents are just function calls now

    The “Recursive Language Model” hands the model one tool — a persistent IPython kernel — and every sub-agent spawns as await rlm('sub-task'), returning at admission instead of blocking. Skills and tools live as Python in that same kernel, so the agent runs functions over its data rather than re-reading it through tool calls.

  • It rewrites its own harness with /refine

    Run /refine and the agent reads its own trajectory, then applies the smallest CRUD edit to its prompt, skills, sub-agents, or memory — every change logged and rollback-able by transaction ID. The base system prompt stays immutable as a guardrail.

  • It edged the human expert on ARC-AGI-3

    Paired with Opus 5, Prime Agent scored 95.5% Best@1 on ARC-AGI-3 — just past the 95.4% human-expert baseline — and 99.97% Best@3 across all 183 levels. On EmulatorBench it wrote working SEGA Genesis and Game Boy Color emulators in Rust from spec, with no reference code.

  • Bring your own login or your own GPU

    It runs on subscription logins you already pay for — Claude Pro/Max, Codex, Copilot — plus API keys (Anthropic, OpenAI, Google, Groq, Fireworks), Bedrock, Azure, and self-hosted vLLM, Ollama, or LM Studio. In their tests GLM-5.2, Opus 5, and GPT-5.6 Sol each beat their native harnesses on 6-8 of 9 long-context benchmarks.

  • Not a sandbox — isolate it before you run it

    Prime Intellect is explicit that worker and kernel processes are not a security boundary; sub-agents executing arbitrary Python in a shared kernel means a throwaway CI box, not your laptop. Build this weekend: point it at a repo behind a test gate and let it refactor overnight.