Prime Agent

Overview

Prime Agent is an open-source self-improving reinforcement-learning agent shipped by Prime Intellect on Aug 5, 2026. Trained to update its own policy from execution traces rather than relying purely on prompt scaffolds; reports 95.5% on ARC-AGI-3 at launch. Fits the 2026 pattern of RL-trained agents replacing prompt-heavy orchestration harnesses.

Timeline

2026-08-06-AI-Digest — Public launch. HN discussion 138 pts / 23 cmts; ARC-AGI-3 headline result 95.5%. Continues the 2026 trend of agents that update their own policy from execution traces.
2026-08-25-AI-Digest — Prime Agent surfaces today via arXiv:2608.23552 (▲18,200) as an open-source Recursive Language Model framework with a persistent IPython REPL, a Continual Harness carrying histories/memories/skills across runs, and recursive subagents that communicate directly (inspectable via an Agents View). The paper reports ARC-AGI-3 RHAE Best@1 lifting 30% → 95.5% plus gains on long-context coding, GPU-kernel generation, emulator construction, nanoGPT speedruns, and parallelised Factorio play. Narrow read the digest carries: openly available harness that separates strategy (model-chosen) from execution / verification / accounting — the corpus baseline for anyone building persistent coding agents just moved. Structural read: pairs with Apodex 1.1 (Environment + Coordination Scaling) and NVIDIA AVO (Opus 5 30→100 on ARC-AGI-3 public set) as the fourth “harness-first” beat this month — the frame is accumulating cross-lab, but at risk of becoming a monoframe until harness components can be measured independently of the base model’s reasoning config.

Key Developments

Launch (Aug 5, 2026): Open-source; self-improving via execution-trace RL; 95.5% ARC-AGI-3.