Prime Agent is a self-improving coding and research agent built for long-running autonomous work — an agent harness whose context, skills, and operating patterns keep getting better the longer it runs.
Prime Agent is an open-source project by Prime Intellect, hosted on GitHub with around 22,000 stars and daily activity in 2026. It is designed around two ideas. The Recursive Language Model treats context as variables and tools as function calls inside a persistent Python REPL, so the agent works programmatically rather than through one-shot prompts. The Continual Harness stores supplemental prompts, memories, and reusable subagent specifications as durable state; a /refine command reviews a session’s trajectory and applies small, evidence-backed improvements, with snapshots for rollback. Daemon-backed sessions keep running after the terminal disconnects.
Prime Agent grows out of Prime Intellect’s research stack — the same organisation behind the Verifiers library and PRIME-RL — and is explicitly aimed at evaluation and research workloads as well as everyday coding. Practical touches include Slack integration through MCP, clipboard forwarding over SSH and tmux, and support for custom certificate authorities, all aimed at running unattended for long stretches inside real engineering environments.
Prime Agent is free and fully open source under the MIT license. Install it with a single command and log in with your own model subscription or API keys; you only pay your model provider.
Researchers and power users running long, autonomous coding and evaluation tasks — people who need an agent that keeps context for days, improves its own working patterns, and can be inspected and steered while it runs.