Argus: New Agentic Runtime for Long-Horizon AI Reasoning

Long-horizon reasoning in AI requires an agentic runtime capable of adapting, persisting through challenges, and pivoting when faced with failures or new constraints. Researchers have introduced Argus, a novel, self-evolving runtime designed to meet these demands. Argus operates with distinct roles—Manager, Planner, Engineer, and Reviewer—each executing bounded missions and maintaining a durable project state. A key innovation of Argus is its ability to self-evolve not through model weight changes, but via persistent runtime state and control policy, with autonomous execution between operator-defined escalation points. It rigorously reviews and verifies memories, skills, and procedures before adoption. Benchmarking shows Argus significantly outperforms Direct Copilot on SWE-Bench Pro, achieving higher success rates with comparable token usage. Furthermore, its self-evolution leads to reduced token consumption and active workflow time in mature stages, demonstrating its efficiency and robustness in complex tasks like code generation, mathematical data synthesis, and scientific paper pipelines.