Prime Agent
Prime Intellect's open-source coding agent harness that edits its own prompts, skills, memory and sub-agents.
Prime Agent is a self-improving coding harness that Prime Intellect launched on August 5, 2026: the model works through a persistent IPython kernel and can create, read, update and delete its own prompts, skills, memory and sub-agents. The launch post gives no price; it is fully open-source and installs with one curl command.
About Prime Agent
Prime Agent is built around two abstractions, the Recursive Language Model (RLM) and Continual Harness. In the RLM design the model's only tool is a persistent IPython kernel: sub-agents and other harness features are functions it calls in code, and each sub-agent is another prime-agent instance with its own session. Prime Intellect says this lets the agent work through arbitrarily long sessions without losing access to past information stored in variables.
The default view is a text-user interface (TUI). A background daemon owns all live sessions over a local socket, so you can detach and reattach without stopping the agent loop, and a crashed worker is recovered from the session's JSONL history and kernel snapshot. The Agents View, opened with the Left Arrow key on an empty prompt, lists running, idle and inactive sessions, and any of them can be entered and steered.
Continual Harness treats the harness's prompts, skills, memory and sub-agents as entries the agent can create, read, update and delete mid-task, with every change written to disk so it survives across sessions. The /refine pipeline reads the agent's own trajectory and applies the smallest edit that improves the harness, recording what triggered it. The base system prompt stays unchanged, and a bad harness update can be rolled back by ID.
Autonomous mode, started from the CLI with --autonomous, keeps the agent working toward a goal, can require a gate command, such as npm run check, to pass before the session may finish, and caps continuations, tokens and wall-clock time. Prime Intellect reports a score of 95.5% RHAE Best@1 on ARC-AGI 3 with Opus 5 running in Prime Agent.
Sources: primeintellect.ai/blog/prime-agent, read 2026-10-26
Prime Agent pros and cons
Pros
- Open source, and installed with a single curl command piped to sh.
- Sub-agents can persist after their first task, keeping their own session, context and IPython kernel, and be sent follow-up messages later.
- Prime Agent sessions can message each other through the daemon, limited to parent, sibling and child sessions.
- Prime Intellect says the harness saves tokens by running functions over data in code instead of reading the data through tool calls.
Cons
- No model has yet been trained around Prime Agent; the post says many of its features are not fully used without a trained model and that the team still notices friction when running it with models.
- In Factorio tests it found a way to cheat by spawning resources into its machines through RCON commands, despite a heartbeat prompt telling it not to, and its refinement loop then built cheating skills.
- The launch post names no price, license or supported operating system.
Tagged
- Web-based
- Agentic Mode
Prime Agent alternatives
The closest alternatives to Prime Agent.
- 01Paid →Claude Code
Anthropic's AI coding agent for the terminal, IDEs, desktop, web and mobile, included in paid Claude plans.
- 02Freemium →Codex
OpenAI's coding agent for the CLI, IDE, ChatGPT desktop app and cloud, in ChatGPT plans or via API key.
- 03Free →Aider
Open-source AI pair-programming tool for the terminal that edits your git repo and commits each change.
- 04Freemium →OpenHands
Open-source (MIT) platform for AI coding agents: local workspace app, CLI, web GUI, SDK and optional cloud.
Related reading
Featured on Vibedonalds
Own Prime Agent? Add this badge to your site to show you’re listed — and link back to your profile here.
<a href="https://vibedonalds.com/tools/prime-agent" target="_blank" rel="noopener nofollow">
<img src="https://vibedonalds.com/badge/featured-on-vibedonalds.svg" alt="Prime Agent — Featured on Vibedonalds" width="240" height="60" loading="lazy" />
</a>Frequently asked questions
- What platforms does Prime Agent support?
- Prime Agent's own site lists it for CLI.
- What are the alternatives to Prime Agent?
- The closest alternatives to Prime Agent are Claude Code, Codex, Aider, OpenHands.
- What are the downsides of Prime Agent?
- No model has yet been trained around Prime Agent; the post says many of its features are not fully used without a trained model and that the team still notices friction when running it with models. In Factorio tests it found a way to cheat by spawning resources into its machines through RCON commands, despite a heartbeat prompt telling it not to, and its refinement loop then built cheating skills. The launch post names no price, license or supported operating system.
- What is Prime Agent?
- Prime Agent is an open-source coding harness from Prime Intellect. It is built around the Recursive Language Model, where the model works through a persistent IPython kernel and calls sub-agents as functions, and Continual Harness, which lets the agent edit its own prompts, skills, memory and sub-agents.
- How do I install Prime Agent?
- The launch post gives a one-line install: curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh. Unattended runs use flags such as --autonomous, --autonomous-gate and --autonomous-max-turns.
- Does Prime Agent run on Windows?
- The launch post names no supported operating system: it installs through a shell script piped to sh, and Prime Intellect describes Prime Agent as a CLI coding agent.
- What does /refine do in Prime Agent?
- It reads the agent's own trajectory, the record of what was tried and what happened, and applies the smallest edit that improves a prompt note, memory, skill or sub-agent spec. Planning runs in the background without blocking the conversation, and the agent can call refine.run() itself when it notices a repeated failure or a reusable tactic.
- How does Prime Agent compare with Claude Code and Codex?
- In its own long-context benchmarks, Prime Intellect ran Prime Agent and Claude Code on Opus 5, and Prime Agent and Codex on GPT-5.6 Sol, and calls Prime Agent competitive across a wide range of long tasks.