English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Argus: A Self-Evolving Agent Runtime for Long-Horizon Reasoning (arXiv:2608.05144)

Forum topic · 小凯 · 2026-08-13

Summary

Argus is a general-purpose agentic reasoning runtime for long-horizon tasks, proposed by researchers from Shanghai Jiao Tong University, Microsoft, Fudan University, and Tsinghua University (arXiv:2608.05144). It addresses goal drift and loop-stalling in agents by making the goal itself co-evolve with the problem being solved. Argus uses a four-role architecture (Manager, Planner, Engineer, Reviewer) built around a work contract that separates a fixed standing user intent from an evolving operational objective, constraints, and verification criteria. Model weights are frozen; self-evolution happens at runtime through memory, skills, verifiers, and routing policies that enter the store only after role review and task-native verification. Reported results include ~78% on SWE-Bench Pro (vs 59% for Direct Copilot), 76.8% on AARRI-Bench, +28.0 on math data synthesis, and six autonomously completed paper-production pipelines spanning 640 campaign hours with all pipelines reaching submission completion. Notably, seven no-go rollbacks reframed a failed method search into a 4500-line negative-results study, and an optimized RWKV6 kernel was merged upstream.

Paper: Argus: A General-Purpose Agentic Reasoning Runtime for Long-Horizon Tasks arXiv: 2608.05144 Authors: Boxiu Li, Zimo Wen, Yijia Fan, et al. (Shanghai Jiao Tong University, Microsoft, Fudan University, Tsinghua University, and others)

The Core Problem

When today's agents tackle long-horizon tasks, the goal is often unclear from the start. As they run, they either "play dumb" (goal drift) or fall into dead loops. Argus targets scenarios where the goal itself must co-evolve with the problem being solved — not exam questions, but genuine research.

Four-Role Architecture

| Role | Responsibility | |------|----------------| | Manager | Anchors user intent; holds authority over work-contract changes | | Planner | Decomposes the next task unit | | Engineer | Executes, implements, evaluates | | Reviewer | Audits artifacts and issues completion verdicts |

Key Design — Work Contract Kt = (ι, ot, ct, vt)

  • ι (standing user intent): stable user intent — immutable
  • ot (operational objective): the current actionable goal — revised based on evidence
  • ct (constraints): known constraints — updated with discoveries
  • vt (verification criteria): verification standards — adjusted as understanding deepens
  • This separation lets the agent revise its operational path based on evidence without tampering with the original intent. The Manager holds change authority, but every change must be backed by evidence.

    Verification-Gated Self-Evolution

    Model weights do not change. Evolution happens in runtime state — memory, skills, verifiers, and routing policies — and only enters the store after role review and task-native verification.

  • Candidate generation → accountability checks → authorized commit
  • Failed branches can become reusable progress
  • Unverified narratives are never allowed to rewrite the entire task
  • Results

    | Benchmark | Score | Comparison | |------|------|------------| | SWE-Bench Pro | ~78% | Direct Copilot 59% | | AARRI-Bench | 76.8% | — | | Math data synthesis | +28.0 pts | — |

  • Token consumption is 1.41x, but mature waves save 21% solve tokens and 15% active time vs startup waves
  • Records 34 verifier recoveries and 22 strict-review loop rescues
  • An optimized RWKV6 kernel was merged upstream
  • Paper-Pipeline Case Study

    Six projects autonomously completed paper-production pipelines:

  • 640 campaign hours
  • 254 bounded tasks
  • 576 Engineer turns
  • 286 Reviewer revisions
  • 16 stage rollbacks
  • All 6 pipelines reached submission completion
A representative case: 7 no-go rollbacks reframed a failed method search into a 4500-line negative-results study, then fixed two late-stage commit defects.

One-Line Summary

Argus doesn't make agents smarter — it makes them not silently change goals, not fake completion, and hand unsolvable problems back to humans in the fog of long-horizon work.

---

Reference: https://arxiv.org/abs/2608.05144

Tags

#ai-agents#long-horizon-tasks#argus#runtime#self-evolution#swe-bench#paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633421