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

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Boxiu Li, Zimo Wen, Yijia Fan, Junxiang Lei, Sufeng Guo, Jiaao Wu, Ruize Tang, Mukai Li, Yifei Shen, Xiaoyu Chen, Wanbo Zhang, Runjing Gu, Yifei Gao, Yuheng Wu, Xuyao Huang, Zelong Zhao, Jiachen Zhang, Shibo Hu, Hangxi Guo, Yilin Chen, Yuzhe Zhang, Fan Yang, Chuan Wen, Xian Zhang, Xuanhe Zhou, Zhijie Deng
• Venue: arXiv: cs.AI
• Published: 2026-08-05
• Collected: 2026-08-06
• URL: http://arxiv.org/abs/2608.05144v1
• PDF: https://arxiv.org/pdf/2608.05144v1
• DOI: UNKNOWN

Plain-Language Brief

The paper introduces Argus, a persistent and self-evolving runtime system designed for long-term reasoning. This system separates stable user intent from operational objectives and allows for the addition of various components after review and verification. Across seven benchmarks, Argus outperformed another system called Direct Copilot by achieving higher scores, using fewer tokens, and taking less time per task.

What the Document Says

• Argus is a persistent, self-evolving runtime for long-term reasoning.
• It consists of Manager, Planner, Engineer, and Reviewer modules that execute bounded missions over durable project state.
• Argus separates stable user intent from operational objectives, constraints, and verification criteria.
• It allows memories, skills, procedures, verifiers, routing decisions, and rejected routes only after role-owned review and task-native verification.
• Model weights remain fixed in Argus, but self-evolution occurs through persistent runtime state and control policy.
• Across seven GPT-5.5 benchmark arenas, Argus achieved about 78% on SWE-Bench Pro compared to 59% for Direct Copilot while using 1.41 times the aggregate tokens.
• After verification-gated self-evolution, mature SWE-Bench waves used 21% fewer solve-input tokens and 15% less active workflow time per task than startup waves.
• Argus also achieved 76.8% on AARRI-Bench and a significant gap on mathematical data synthesis.
• It has competitive results in GPU-kernel and language-model training.
• Beyond benchmarks, Argus demonstrated successful outcomes in various applications, such as merging an optimized RWKV6 kernel upstream, conducting multi-day mathematics campaigns, and completing six paper pipelines with 16 stage rollbacks.

Key Concepts Mentioned

• Argus
• long-horizon reasoning
• GPT-5.5 benchmark arenas
• SWE-Bench Pro
• Direct Copilot
• self-evolving harness

Claims Made by the Document

• Argus outperforms Direct Copilot across seven GPT-5.5 benchmarks.
• After verification-gated self-evolution, mature SWE-Bench waves use fewer tokens and less active workflow time per task than startup waves.
• Argus achieves higher scores on AARRI-Bench and demonstrates significant success in mathematical data synthesis.

Evidence Used

• Results across seven GPT-5.5 benchmark arenas are provided, showing Argus's performance compared to Direct Copilot.
• Data on the reduction of solve-input tokens and active workflow time per task for mature SWE-Bench waves is presented.
• Results showing Argus's performance on AARRI-Bench and mathematical data synthesis are provided.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Long-horizon reasoning requires an agentic runtime that can persist when evidence supports its current approach and pivot when measurements reveal failure, hidden constraints, or a misspecified objective. We present Argus, a persistent, self-evolving runtime in which Manager, Planner, Engineer, and Reviewer execute bounded missions over durable project state. Argus separates stable user intent from operational objectives, constraints, and verification criteria, and admits memories, skills, procedures, verifiers, routing decisions, and rejected routes only after role-owned review and, when available, task-native verification. Model weights remain fixed; self-evolution occurs through persistent runtime state and control policy, with autonomous execution between operator-owned escalation points. Across seven GPT-5.5 benchmark arenas, Argus achieves about 78% on SWE-Bench Pro versus 59% for Direct Copilot while using 1.41 times the aggregate tokens. After verification-gated self-evolution, mature SWE-Bench waves use 21% fewer solve-input tokens and 15% less active workflow time per task than startup waves, while recording 34 verifier recoveries and 22 strict review-loop rescues. Argus also reaches 76.8% on AARRI-Bench and a 28.0-point gap on mathematical data synthesis, with competitive GPU-kernel and language-model-training results. Beyond benchmarks, an optimized RWKV6 kernel was merged upstream; a multi-day mathematics campaign retained falsified routes and proof-backed frontier updates; and six paper pipelines completed 254 missions with 16 stage rollbacks. These results show that a fixed-weight, self-evolving harness can revise, recover, and accumulate verified approaches while producing structured trajectories for future supervised and reinforcement learning.