Prime Intellect released Prime Agent on August 5, 2026, an open-source coding harness that rewrites its own instructions while it works. Paired with Opus 5, it scored 95.5% RHAE Best@1 on ARC-AGI-3, edging past the 95.4% human-expert baseline on a benchmark where raw frontier models still score below 1%. The gain comes from the harness, not a new model, which is why it matters to anyone building with AI.
What Prime Agent Is
Prime Agent is a self-improving agent for coding and long-running autonomous work, shipped under an MIT license on GitHub. It is free to run with your own API keys or a subscription, and it works with both open and closed frontier models.
Two ideas sit at its core. The first is the Recursive Language Model (RLM), which treats the prompt as a variable and tools as function calls inside a persistent Python kernel. Instead of picking from a static tool list, the model writes small programs, spawns sub-agents with rlm(...), and manages its own context as data.
The second is the Continual Harness. Prompts, skills, memories, and sub-agent definitions are stored as durable state the agent can read and rewrite across sessions. In practice, the agent treats its own configuration as something it can edit, not a fixed wrapper it is stuck inside.
How It Beats Fixed Harnesses
The differentiator is the /refine command. It reads the agent's trajectory, then applies minimal create, read, update, and delete edits to the harness state rather than rewriting everything. Planning runs asynchronously in the background, changes apply at turn boundaries, and every refinement is evidence-backed and rollback-able by ID. That is a real departure from the fixed-configuration approach used by tools like Claude Code and Codex.
The number is striking because of where the field started. ARC-AGI-3 is designed to be hard for models and easy for people: raw frontier models score below 1%, with Gemini 3.1 Pro managing 0.37%. Getting from there to above the human line without touching the model weights is the whole point. The lift lives in how the agent organizes its own reasoning, spawns helpers, and prunes what is not working, not in a bigger network.
Prime Intellect reports that on long-context tasks, Prime Agent stays competitive with or ahead of proprietary harnesses while using fewer tokens, and that the improvement holds even with open-weights models such as GLM-5.2. Because no model was trained specifically for Prime Agent, the harness gains transfer across whatever model you point it at. The same open-harness portability drives tools we have covered before, like OpenCodex running any LLM inside a Codex-style loop.
| Dimension | Prime Agent | Claude Code | Codex |
|---|---|---|---|
| ARC-AGI-3 (best result) | 95.5% (Opus 5) | Vendor official | Vendor official |
| Harness state | Self-modifiable via /refine | Fixed | Fixed |
| Models | Any open or closed | Claude only | OpenAI only |
| License | MIT open-source | Proprietary | Proprietary |
| Multi-agent | Persistent, survives disconnect | Limited | Limited |
Run It With Your Own Models
Getting started takes three steps and no vendor lock-in. The documentation walks through provider setup in detail.
1. Install. Run curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh on macOS or Linux.
2. Launch in your project. Change into your repository and run prime-agent. On first launch, run /login to pick a subscription or an API-key provider such as Opus 5, GPT-5.6 Sol, or an open-weights model.
3. Let it learn. Work a task, then run /refine so the agent updates its own prompts, skills, and memory based on what worked. For unattended work, autonomous mode adds token budgets, turn limits, and completion gates via CLI flags.
Why It Matters for Builders
For anyone building apps, sites, or content pipelines with AI, the orchestration layer is where projects stall. A harness that adapts to your codebase and carries persistent, messageable sub-agents across terminal disconnects is closer to a durable coworker than a chat window. It slots into the same shelf as self-hosted agent runtimes like OpenWorker, but adds the self-editing loop on top.
There is also a cost angle. A harness that reaches a target with fewer tokens, and that can run a cheaper open-weights model when a task does not need a frontier one, changes the math on long autonomous jobs where token spend compounds over hours. For a solo builder or a small studio, the difference between a fixed premium-model loop and a tunable open one is the difference between an experiment and a habit.
The open license is the real unlock. You can read exactly how the agent modifies itself, fork it, and swap the model underneath without waiting on a vendor. That portability is what turns a benchmark headline into something a small team can actually run in production this week.
The Caveats
Two honest limits ship with the hype. First, the 95.5% figure is a self-reported result from Prime Intellect's announcement and has not been independently endorsed by the ARC Prize team, so treat it as reported rather than certified. Second, the team disclosed that during Factorio tests the self-improvement loop discovered reward-hacking strategies, a reminder that an agent allowed to rewrite its own goals can optimize for the wrong thing. The RLM approach itself builds on prior research, credited to work published in late 2025 and detailed in an accompanying arXiv paper.
Frequently asked questions
Is Prime Agent free to use?
Yes. Prime Agent is open-source under the MIT license and free to run. You supply your own model access through an API key or a subscription provider when you run /login.
Which models does Prime Agent support?
It works with both open and closed frontier models. Prime Intellect evaluated it with Opus 5, GPT-5.6 Sol, and the open-weights GLM-5.2, and the harness is model-agnostic by design.
What does the /refine command actually do?
It reviews the agent's recent trajectory and applies small, evidence-backed create, read, update, and delete edits to the harness state, its prompts, skills, and memory. Every change is reversible by ID.
How is it different from Claude Code or Codex?
Claude Code and Codex use fixed harnesses tied to one vendor's models. Prime Agent is open-source, runs any model, and can rewrite its own configuration mid-task through the Continual Harness and /refine.
Is the 95.5% ARC-AGI-3 score verified?
Not independently. It is Prime Intellect's own reported RHAE Best@1 result using Opus 5. The ARC Prize team has not certified it, so treat the number as a strong claim pending third-party confirmation.
Can it run long, unattended tasks?
Yes. Autonomous mode supports long-running work with token budgets, turn limits, and completion gates, and its sub-agents persist and stay messageable even after a terminal disconnects.
