Prime Mind has open-sourced Prime Agent, a self-improving coding harness designed round two abstractions, the Recursive Language Mannequin (RLM) and Continuous Harness. Mounted instrument schemas and context compaction drive a mannequin to work round its personal scaffolding. Prime Agent replaces each with a persistent Python REPL and a rewritable harness. With Opus 5, it studies 95.5% on ARC-AGI-3, above the reported human professional baseline of 95.4%. It’s MIT-licensed.
Is it deployable
Sure, in the present day. Prime Agent installs on Linux or macOS with one command. It runs on subscription logins (Codex, Claude Professional/Max, GitHub Copilot), API keys (Anthropic, OpenAI, Google, Groq, Fireworks, Prime Inference, and others), Azure OpenAI, Amazon Bedrock, and self-hosted vLLM, Ollama, or LM Studio endpoints. Self-hosting an open-weights mannequin akin to GLM-5.2 retains code inside your personal community.
- Firm stage: Greatest match is mid-size to massive engineering orgs and AI labs that already run remoted CI containers. Prime Mind states plainly that employee and kernel processes are not a safety sandbox. Deployment due to this fact wants disposable clones or restricted environments. Solo builders can set up it, however the payoff seems on multi-hour duties.
- Industries: Developer tooling, semiconductor and HPC groups writing GPU kernels, simulation and gaming, quantitative analysis, and AI analysis labs.
- Purposes: In a single day refactors behind a check gate, spec-driven builds from scratch, kernel optimization, long-horizon agent analysis, and autoresearch.
What Prime Mind shipped
Prime Agent is constructed on two abstractions. The Recursive Language Mannequin (RLM) treats context as a variable and sub-agent delegation as perform calls inside a REPL. The Continuous Harness treats prompts, sub-agents, expertise, and reminiscence as state the agent can create, learn, replace, and delete from its personal trajectory. Each papers have Prime Agent authors on them. The TUI is constructed on pi.
