Introducing Prime Agent: A self-improving RLM harness for coding and long-running autonomous tasks. Designed to be both token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state.

Prime Agent is a general-purpose coding harness On ARC-AGI-3, it scores 95.5%, surpassing the human-expert baseline, but the gain is not benchmark-specific. We see major improvements across models when compared to their proprietary harnesses:

Prime Agent combines three ideas: 1. Recursive Language Models-native programmatic tool calling 2. Persistent multi-agent orchestration 3. A self-improving Continual Harness Together, they let the model act on its own context and harness

The RLM treats context as a variable. A persistent IPython kernel is the model's only tool. It can program over history, call tools, launch sub-agents, and keep useful state outside the active context. Long sessions become a programming problem.

We recommend using Prime Agent for both day-to-day work and standalone evals. We find strong downstream performance gains for open and closed frontier models using Prime Agent on standard long-context and long-horizon tasks, despite modern models not being trained around Prime Agent.

On EmulatorBench, it built SEGA Genesis and Game Boy Color emulators from scratch in Rust, reproducing target hardware behavior against diagnostic tests.

Prime Agent was designed as a coding agent, but also can be used for general agentic tasks in any domain. Prime Agent can autonomously run on challenging long-horizon settings like Factorio and MazeBench and yield strong results relative to existing harnesses.

Prime Agent is built on pi and fully open-source with an open license

Blog:

Install prime-agent curl -fsSL | sh

Read the full blog

I can’t believe my eyes. At the eod it was a harness problem

huge. been wondering why RLMs have been overlooked in harnesses. Can't wait to try it out

I hope they're already working on Arc AGI 5

@Grad62304977 wow incredible work!

banger of an agent

damn everyone just be dropping stuff today, this is cool, gotta try

Nice!

Codex, download prime intellect on my computer

holy ARC-AGI 3 solved

you forgot to mention in post that you're talking about 25 public games, and btw there are few harnesses that scored more than 96% in them no agi today, unfortunately

Good stuff! I built on the same idea a couple of months back . I hope this succeeds because the core idea is a no-brainer and needs to be implemented and maintained with company backing.

Excellent work!

i love you guys so much man

Everyone on X when Prime Intellect drops Prime Agent

the longer it runs, the more state builds up, and locking all of it into one platform's infrastructure makes leaving expensive later

This looks very interesting! But from first few minutes of use, i'm noticing it lacks some of the basic usability polish that pi has, here's an example:

waowh

If you can make Opus smart again I will name my son “Prime”

This is incredible work.

now time to put it on and

cool futuristic aesthetics! i dig that the ARC-AGI 3 benchmark being on the public set should have been IMO mentioned but its pretty cool tho "/refine is the self-improving pipeline built on top of this CRUD surface. It reads the agent's own trajectory, the record of what was tried and what happened, and applies the smallest relevant CRUD edit that improves the harness toward better outcomes: updating a prompt note, memory, skill, or sub-agent spec, rather than rewriting the whole harness." but i see its not selfimproving the whole harness yet, like darwin godel machine, but still cool also lmao "However, we also observed instances of reward hacking by Prime Agent in FLE. Prime Agent discovered it could bypass Factorio's rules entirely by spawning in resources directly into its assembly machines through RCON commands, even with an explicit heartbeat prompt to remind Prime Agent not to cheat in Factorio. Once it found this exploit, the same refinement loop that had been building legitimate skills turned to building efficient cheating skills instead."

So cool! If only @liquidai had this for their recent LFM2.5 release...

Easy view of tool definition and system prompt:

@thsottiaux I smell fear 😅

insanely good work, congrats!

construct v0.17.4 adds Prime Agent support, and I had Prime Agent itself run the release. it did fine! runs next to my claude code / codex sessions now!

cool work!!

brb finding a way to add this banger harness to box wtf

Why is ARC-AGI-3 even letting harnesses partake ? it should be a single model context test, not a harness test, you are brute forcing your way into fake AGI, this is cheating

Raised a PR to integrate @nebiustf 👋 cc @vincentweisser

Epic release!!!

awesome. great work!! 👏👏

You should see if it can play Rimworld