Prime Intellect Releases Open-Source Prime Agent Coding Harness

Prime Intellect has released Prime Agent, an open-source coding harness designed for self-improvement through two core abstractions: the Recursive Language Model (RLM) and the Continual Harness. This system replaces traditional fixed tool schemas and context compaction with a persistent Python REPL (Read-Eval-Print Loop) and a rewritable harness, allowing models to interact with their environment dynamically. Prime Agent is designed to handle complex coding tasks by treating context as a variable and sub-agent delegation as function calls within the REPL. The Continual Harness manages prompts, sub-agents, skills, and memory as state that the agent can manipulate throughout its operational trajectory. The company reports that Prime Agent, utilizing Opus 5, achieved a 95.5% score on the ARC-AGI-3 benchmark, surpassing the reported human expert baseline of 95.4%. The software is distributed under the MIT license and is available for immediate deployment on Linux or macOS with a single command. Prime Agent supports various operational modes, including subscription logins for services like Codex, Claude Pro/Max, and GitHub Copilot, as well as API keys from providers such as Anthropic, OpenAI, Google, and Groq. It also integrates with cloud platforms like Azure OpenAI and Amazon Bedrock, and can be self-hosted using vLLM, Ollama, or LM Studio endpoints. For organizations prioritizing data privacy, self-hosting an open-weights model like GLM-5.2 allows code to remain within the company's network. Prime Intellect recommends Prime Agent for mid-size to large engineering organizations and AI labs that already utilize isolated CI (Continuous Integration) container environments. The company explicitly states that worker and kernel processes within Prime Agent do not constitute a security sandbox, necessitating deployment in disposable clones or restricted environments. While solo developers can install Prime Agent, its full benefits are most apparent on tasks requiring multiple hours of operation. The target industries for Prime Agent include developer tooling, semiconductor and HPC teams developing GPU kernels, simulation and gaming companies, quantitative research firms, and AI research laboratories. Potential applications range from overnight code refactoring behind test gates and spec-driven development from scratch to kernel optimization, long-horizon agent evaluation, and automated research processes. The TUI (Text User Interface) for Prime Agent is built on the pi library.