Overview
- Prime Intellect publicly unveiled Prime Agent on Wednesday, presenting it as open-source infrastructure that runs long-lived agent sessions and lets agents modify their own harness state during a run.
- The system’s core mechanism is a 'Continual Harness' that runs inside a persistent Python REPL so context persists as programmatic variables and the agent can invoke tools, spawn sub-agents, inspect files, and run tests.
- Prime Intellect reported that Prime Agent using Opus 5 scored 95.5% on the ARC-AGI-3 benchmark and said that result exceeds the benchmark’s human‑expert baseline, with the coverage noting that independent verification of the claim was not provided.
- The company released supporting components including a verifiers stack (v1) and an Environment Hub with hundreds of thousands of sandboxed environments, and it said Prime Agent will be free of licensing fees or vendor lock-in after a $130 million Series A and a reported $1 billion valuation.
- The launch builds on Alex Zhang’s Recursive Language Model research and could shift value toward model‑harness co‑design and infrastructure, but it raises questions about safety, reproducibility, and how models and harnesses should be trained together.