Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning

За 30 секунд

Что произошло

Argus is an agentic runtime system that manages long-horizon reasoning tasks through persistent state and self-evolution, achieving 78% on SWE-Bench Pro versus 59% for Direct Copilot.

Почему это важно

Software engineers building AI agent systems should care when they need reliable multi-step task execution with recovery from failures and the ability to learn from experience without retraining models.

На что обратить внимание

Argus uses 1.41 times more aggregate tokens than baselines and requires operator-owned escalation points, meaning it trades computational efficiency for reliability and control in complex reasoning tasks.