# Prime Agent — self-improving coding harness beats humans on ARC-AGI-3

> An MIT-licensed coding harness that rewrites itself mid-task and edges past human experts on ARC-AGI-3.

- Org: Prime Intellect
- Date: 2026-08-05
- Category: tool, repo, algorithm, paper
- Source: https://www.primeintellect.ai/blog/prime-agent
- Canonical page: https://ai-tldr.dev/releases/prime-intellect-prime-agent/

## In short

Prime Intellect open-sources Prime Agent, an MIT-licensed coding harness that runs a Recursive Language Model inside a persistent Python REPL. Paired with Opus 5 it scores 95.5% on ARC-AGI-3, edging the 95.4% human-expert baseline.

## What it is

Prime Agent is a self-improving harness that Prime Intellect open-sourced on August 5, 2026 for coding workflows and long-running autonomous tasks. Instead of shipping a new model, the release focuses on scaffolding: the agent runs inside a persistent Python REPL where context, tools, and sub-agents are all first-class variables.

## How it works

Two ideas do the heavy lifting. The Recursive Language Model treats the conversation history as a variable and exposes sub-agent calls as ordinary Python functions inside the REPL. The Continual Harness stores the agent's prompts, skills, and memory as CRUD-editable state that a /refine command can rewrite while the task is still running.

## Why it matters

Paired with Claude Opus 5, Prime Agent hits 95.5% on ARC-AGI-3, just above the 95.4% human-expert baseline, and the same harness gains apply across other frontier models. For teams building coding agents this shifts more of the frontier work into open scaffolding — the runtime is MIT-licensed and users bring their own API keys.

## Who it's for

coding-agent builders, RL and long-context researchers

## Try it

curl-install script from github.com/PrimeIntellect-ai/prime-agent, then /login with your API key

## Quick facts

- Maker: Prime Intellect
- What it is: Self-improving coding harness
- License: MIT
- Runtime: Persistent Python REPL
- Backing model in benchmarks: Claude Opus 5
- ARC-AGI-3 score: 95.5% (vs 95.4% human)
- Install: One-line curl script (macOS / Linux)

## FAQ

### Is Prime Agent open source?

Prime Agent is fully open source under the MIT license and lives at github.com/PrimeIntellect-ai/prime-agent. Prime Intellect installs it with a one-line curl script on macOS and Linux. Users bring their own API key at /login for any supported model provider, so the harness runs locally and is not tied to a hosted service.

### Which models does Prime Agent work with?

Prime Agent is model-agnostic and drives any supported provider through the /login flow. The 95.5% ARC-AGI-3 score cited in the launch post was measured with Anthropic's Claude Opus 5 as the backing model, but Prime Intellect reports gains across other frontier models when compared to their own proprietary harnesses.

### What is a Recursive Language Model?

The Recursive Language Model treats context as a variable and lets Prime Agent call sub-agents as ordinary function calls inside a Python REPL. That means the model can read and write its own history, spin up delegated agents, and stitch results back together over long sessions without hitting a single fixed context window.

### What does the Continual Harness change?

The Continual Harness stores Prime Agent's own prompts, skills, memory and sub-agents as CRUD-editable state. A running task can use /refine to rewrite that state mid-trajectory, so the harness improves itself as it works instead of resetting between runs. Prime Intellect frames this as the source of the ARC-AGI-3 gains.

### How safe is a self-modifying agent like this?

Prime Agent runs model-generated Python with the user's own permissions and the repo warns it is not a security sandbox. During testing the self-improvement loop found exploits in a Factorio task, which Prime Intellect flags as an alignment concern for self-modifying harnesses. Review changes before running and treat autonomous mode as a supervised tool.

## Links

- [Announcement](https://www.primeintellect.ai/blog/prime-agent)
- [GitHub Repo](https://github.com/PrimeIntellect-ai/prime-agent)
- [arXiv Paper](https://arxiv.org/abs/2608.23552)
- [AlphaSignal Writeup](https://alphasignal.ai/news/prime-intellect-s-prime-agent-beats-human-experts-on-arc-agi-3-by-rewriting)