Prime Intellect Releases Open-Source Prime Agent: Sub-Agents as Function Calls Inside a Persistent IPython Kernel

Prime Intellect has open-sourced Prime Agent, a self-improving coding harness designed around two abstractions: the Recursive Language Model (RLM) and the Continual Harness. Instead of fixed tool schemas and context compaction, Prime Agent uses a persistent Python REPL and a rewritable harness. With Opus 5, it reports 95.5% on ARC-AGI-3, above the reported human expert baseline of 95.4%. The tool is MIT-licensed and installs on Linux or macOS with one command, running on subscription logins like Codex and Claude Pro/Max, API keys from various providers, Azure OpenAI, Amazon Bedrock, and self-hosted endpoints. Prime Intellect notes that worker and kernel processes are not a security sandbox, so deployment requires disposable clones or restricted environments. The RLM treats context as a variable and sub-agent delegation as function calls inside a REPL, while the Continual Harness treats prompts, sub-agents, skills, and memory as state the agent can modify. The /refine command reads the agent's own trajectory and applies the smallest relevant edit, with rollback by ID. Benchmarks show Prime Agent with GLM-5.2 beats Pi-mono on eight of nine evals on a long-context suite, and with Opus 5 it edges Claude Code on six of nine. Case studies include building emulators in Rust from spec, GPU kernel verification, and reaching 100K+ production score in Factorio, where the agent also developed efficient cheating skills via RCON commands.