Prime Agent
TL;DR
Prime Agent is an open-source coding and research agent framework from Prime Intellect that uses a Recursive Language Model (RLM) architecture, allowing the agent harness to rewrite its own prompts, memory, and tests at runtime for self-improvement. It demonstrates that test-time compute scaling via self-modifying agent infrastructure — not model improvements alone — can close the gap to human-level benchmark performance, boosting Opus 5 from 30% to 95.5% on ARC-AGI-3.
What it actually is
- What: Prime Agent is an open-source coding and research agent framework from Prime Intellect that uses a Recursive Language Model (RLM) architecture, allowing the agent harness to rewrite its own prompts, memory, and tests at runtime for self-improvement.
- Who built it / maintained by: Prime Intellect
- Status: alpha
- Why it matters: It demonstrates that test-time compute scaling via self-modifying agent infrastructure — not model improvements alone — can close the gap to human-level benchmark performance, boosting Opus 5 from 30% to 95.5% on ARC-AGI-3.
- How it compares to alternatives:
- SWE-agent
- OpenDevin
- Aider
- Cursor
- AutoGPT
- Devin
- GitHub stars: 4,552 · License: MIT · Archived: no
Links
Kickstarter guide
Clone the repository at github.com/PrimeIntellect-ai/prime-agent and follow the setup instructions. The framework is built in TypeScript and centers on two abstractions: the Recursive Language Model (RLM), which treats context as a variable and tool calls as program calls, and a Continual Memory system that persists prompts and reusable subroutines across sessions. Point it at your coding task and the harness will self-modify its scaffolding based on observed failures during the run.