Argus: A Persistent Agentic Runtime for Self-Evolution Without Weight Updates

imjustnewatai · x · 2026-08-07

To tackle long-horizon reasoning, a new paper introduces Argus, a persistent, self-evolving agentic runtime where Manager, Planner, Engineer, and Reviewer execute bounded missions over durable project state.

Crucially, model weights remain completely fixed; self-evolution occurs strictly through persistent runtime state and control policy. Memories, skills, procedures, and rejected routes are only admitted after role-owned review and task-native verification. On SWE-Bench Pro, Argus achieves about 78% versus 59% for Direct Copilot. Following verification-gated self-evolution, mature waves use 21% fewer input tokens and 15% less active time per task than startup waves.

Related event: Argus Agent Runtime Enables Self-Evolution Without Retraining(2 posts)→

Original post →

More from coding & agent

coding & agent channel →