Argus: Self-Evolving Agentic Runtime Achieves 78% on SWE-Bench Without Model Retraining

imjustnewatai · x · 2026-08-07

Recursive self-improvement could begin before a model ever retrains. A newly released paper introduces Argus, a persistent agentic runtime that improves while keeping GPT-5.5 weights completely frozen.

Argus assigns distinct Manager, Planner, Engineer, and Reviewer roles. It only saves memories, skills, and failed routes as reusable assets after strict evidence checks. Across 731 software tasks, Argus scored about 78% compared to 59% for Direct Copilot, though it used 1.41x more tokens overall.

Crucially, as verified experience accumulated, mature execution waves used 21% fewer input tokens and 15% less active time per task than startup waves. This demonstrates that the runtime can evolve its control policy around the model, generating high-quality trajectories to eventually train successor models.

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

Original post →

More from coding & agent

coding & agent channel →