Recursive Language Models

Core Event Prime Agent is a newly released open-source framework designed for general-purpose and long-horizon coding and research tasks. It has achieved a landmark 95.5% score on the ARC-AGI-3 benchmark, effectively outperforming human expert baselines and established industry tools like Codex and Codium. ▶ Architectural Paradigm Shift: Transitions from static prompting to a Recursive Language Model (RLM) framework, utilizing programmatic tool calls and state management. ▶ Token Parsimony: Implements "Context Variabilization" to maintain high expressivity while drastically cutting token overhead in complex reasoning chains. ▶ Self-Modifying Autonomy: Features self-modifying states and multi-agent communication protocols, enabling robust performance in multi-step, autonomous problem-solving. Bagua Insight At Bagua Intelligence, we view Prime Agent as a pivotal step toward the "Agentic OS" era. The industry is moving beyond the "scaling laws" of raw parameters; the new frontier is sophisticated orchestration. By treating context as programmable variables rather than a linear stream of text, Prime Agent mitigates the "lost in the middle" phenomenon and information decay. The 95.5% ARC score is a shot across the bow for proprietary labs, proving that architectural innovation in agentic harnesses can leapfrog raw model power in high-stakes logical reasoning. Actionable Advice Developers should pivot from static Prompt Engineering to designing stateful Agentic Workflows, leveraging RLM-style recursive logic. For enterprises, Prime Agent serves as a blueprint for high-efficiency R&D tools—prioritize architectures that support recursive self-correction to handle the inherent complexity of evolving, large-scale codebases.