Prime Intellect's Prime Agent Hits 95.5% on ARC-AGI-3 by Rebuilding AI Scaffolding

Language models are strong sequential processors, but they hit a wall when tasks stretch across hours or days. Context windows fill up. State gets dropped. Tool schemas become straitjackets. Prime Intellect's new release, Prime Agent, tries to fix that not by training a smarter model, but by rebuilding the scaffolding around it. The open-source, self-improving coding harness is built around two abstractions, the Recursive Language Model (RLM) and the Continual Harness, and with Opus 5 it achieves 95.5% on ARC-AGI-3, surpassing the reported human expert baseline.

Why harnesses became the bottleneck

Most agent harnesses today were designed for the previous generation of models. Fixed tool-calling schemas and context compaction force the model to work around its own scaffolding instead of leveraging it, and static, hand-engineered sub-agents, prompts, skills, and memory are set once at design time and never adapt to what the agent learns while running.

The Prime Agent team frames this as an evaluation problem too. When a strong model scores poorly on a benchmark, you often cannot tell whether the model failed or the harness dropped state, restricted useful actions, miscounted resources, or terminated the run too early. The team argues that the harness should be a low-friction membrane so measured performance actually reflects the model's ceiling.

Context as a variable

Models in Prime Agent use a persistent IPython kernel as their only tool, and other standard harness features are called as functions in the kernel, including sub-agents, which are each implemented as another prime-agent instance.

That changes the semantics of a turn. The harness no longer needs to squeeze every file, message and result into the model's immediate context window because large inputs can live in Python variables. The model writes Python to search, filter, summarize, and aggregate those variables, pulling only what it needs into the token stream.