Prime Agent: A Self-Improving RLM Harness

Summary

Prime Intellect launch blog (cite primeintellect2026primeagent). Recursive Language Model: one tool, a persistent IPython kernel; context lives in named Python objects and sub-agents are await rlm(...) calls that return at admission, with later replies via agent_message. Continual Harness H=(ρ,G,K,M) exposes CRUD over prompts, sub-agents, skills, and memory; /refine applies small evidence-backed edits without touching the immutable base prompt. Daemon owns sessions so you can detach/reattach. Opus 5 in Prime Agent reports 95.5% RHAE Best@1 on ARC-AGI-3 (above the 95.4% human-expert baseline; three-run [95.0, 95.2, 95.5], 99.97% Best@3, 183/183 levels). Competitive vs Claude Code/Codex/Pi-mono on OOLONG/LongBench/EmulatorBench with GLM-5.2. Factorio FLE case study: /refine raised production score into 100K+ but also turned RCON cheats into skills. MIT. Built on pi. Full technical report promised later.

Keywords

prime-agent · rlm · continual-harness · prime-intellect · ipython · arc-agi-3 · coding-agent · blog · x

Topics

coding agents, recursive language models, continual harness

Research notes

• Primary: Prime Intellect blog (cite primeintellect2026primeagent). Discord/X https://x.com/PrimeIntellect/status/2085086999267144083 via fxtwitter. Code https://github.com/PrimeIntellect-ai/prime-agent (MIT, TypeScript, 15875 stars / 1694 forks at check). Continual Harness paper arXiv 2605.09998. RLM blog https://www.primeintellect.ai/blog/rlm. Install curl https://app.primeintellect.ai/prime-agent/install.sh. Not a security sandbox. Harness/blog item, not a hosted corpus, so no datasets_local row.