Prime Agent is an open-source harness for long-horizon evaluation and coding-agent workflows. What is new is not a stronger model but a membrane: a persistent IPython REPL and a Continual Harness that carry histories, memories, skills, prompts, and subagent specifications across trajectories, turning context into programmatically managed external state. The abstract reports ARC-AGI-3 RHAE Best@1 rising from 30% to 95.5% with model class held fixed, plus parity or wins on long-context coding, GPU-kernel generation, emulator construction, and nanoGPT speedruns.12
Context
In most agent runs, context is re-derived from a flat transcript and everything except files on disk resets, so a system cannot compound. Prime Agent removes that constraint: the model writes code to process its context in a persistent REPL, and the harness stores the residue of past runs. That is the enabling move; the reported gains are plausibly attributed to persistence rather than model scale.12
Assessment
The evidence is a single self-report. The arXiv abstract and the newsletter repeat the same numbers; the newsletter adds no independent measurements. Neither text names the baseline harness, the model used in the ARC-AGI-3 runs, or the compute budget, so the 65.5-point gain cannot be audited from these documents.12
Caveats
What would change my mind is a reproducible public evaluation: the code URL, the exact RHAE protocol, and a comparison set that shows Prime Agent's harness rather than extra compute is responsible for the improvement. Until then, the headline ARC-AGI-3 number is best read as a claim about a specific benchmark protocol, not general agent capability.12
