Prime Agent Scores 95.5% on ARC‑AGI‑3: A Self‑Evolving AI Agent Framework

AI developers often lose context when a terminal disconnects, and traditional agent frameworks with fixed prompts and tool formats collapse on unfamiliar abstract reasoning tasks.

Prime Agent, an open‑source framework built on Opus 5, addresses these issues and records a 95.5% score (Best@1) on the ARC‑AGI‑3 benchmark, slightly exceeding the human expert baseline of 95.4% and achieving a near‑perfect Best@3 average of 99.97%. In contrast, the native Opus 5 framework scores only 30.2%, while Prime Agent combined with GPT‑5.6 Sol reaches 78.3%, demonstrating the impact of the framework’s enhancements. Additional experiments show strong performance on long‑context OOLONG 128K (0.94 vs 0.50), EmulatorBench (0.998 vs failure), Factorio factory simulations (10 000+ score), 3D MazeBench (far‑superior resource collection), and GPU kernel development.

The core innovation is the Recursive Language Model (RLM). RLM turns the entire conversation into a persistent IPython environment where context, history, and task records become manipulable Python variables. The model’s sole native tool is a background IPython REPL kernel; sub‑agents invoke rlm() asynchronously, enabling parallel task distribution and message exchange. RLM’s default MAX_DEPTH=1 limits recursion for cost control but can be reconfigured, and the design aligns with similar programming‑style tool abstractions used by Codex and Claude Code.

The second breakthrough, Continual Harness, makes prompts, memory, reusable skills, and sub‑agent configurations fully CRUD‑able at runtime. The /refine command implements self‑evolution through four steps: (1) automatically record the full task trajectory, including failures and successful strategies; (2) distill high‑frequency effective solutions into memory fragments that can be imported as Python skills or sub‑agent configurations; (3) enforce permission isolation by keeping base system prompts immutable; and (4) create version snapshots for one‑click rollback of any refinement.

Prime Agent shines in scenarios such as large‑scale code refactoring, algorithmic simulation, GPU kernel tuning, long‑context document analysis, and AI research that requires custom harness development. It is less suitable for tiny one‑off tasks, handling sensitive financial or personal data (due to lack of sandboxing), or projects without automated testing, where the self‑evolution mechanism may generate risky code.

Three critical risks are highlighted: reward‑hacking, where the agent exploits loose evaluation rules to produce non‑compliant outputs; the default recursion depth that can bottleneck complex multi‑layer task decomposition unless manually increased; and uncontrolled resource consumption in autonomous mode, which can lead to excessive API costs if token or time limits are not set.

The framework sparked immediate debate. Critic Peter Wang questioned RLM’s recursion depth and alleged overfitting on ARC‑AGI‑3, while author Alex Zhang clarified that RLM’s essence is “context variableization and tool code‑ification,” that depth is a configurable cost‑control parameter, and that the paradigm’s value extends beyond the benchmark. Both sides agree that further private test sets are needed for definitive validation.

Installation requires macOS or Linux, a one‑click script (

curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh

), login to configure the large‑model API key, and use of the provided CLI commands such as prime-agent agents, prime-agent attach <session‑id>, prime-agent --resume <path>, prime-agent status, prime-agent doctor --fix, prime-agent update --force, and prime-agent shutdown --force.

In summary, Prime Agent demonstrates that harness architecture—not just model size—can unlock substantial reasoning gains, but its benchmark scores should be interpreted cautiously, and its self‑evolution features demand careful guardrails and testing before production deployment.