Argus: A Self-Evolving Agentic Runtime for Long-Horizon Reasoning

Forum topic · 小凯 · 2026-08-13

Summary

Argus (arXiv:2608.05144) is a general-purpose agentic runtime designed for long-horizon tasks where user intent and the problem itself must co-evolve. Co-authored by researchers from Shanghai Jiao Tong University, Microsoft, Fudan, and Tsinghua, Argus introduces a four-role architecture (Manager, Planner, Engineer, Reviewer) coordinated through a mutable Work Contract K_t = (ι, o_t, c_t, v_t) that separates immovable user intent from operational objectives, constraints, and verification criteria. Model weights remain frozen; evolution occurs in runtime state (memory, skills, verifiers, routing) only after role review and task-native verification. On SWE-Bench Pro Argus reaches roughly 78% versus 59% for a direct Copilot baseline, and reports 76.8% on AARRI-Bench and +28.0 points on math data synthesis. Maturation waves reduce solve tokens by 21% and active time by 15%, with dozens of verifier recoveries and rollback rescues. A six-project paper-production pipeline ran 640 campaign hours, 254 bounded tasks, and 576 engineer turns, with all six pipelines reaching submission-complete state.

Paper

Title: 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)

Core Problem

Current agents tackling long-horizon tasks often suffer from unclear goals that drift mid-run, either "playing dumb" (goal drift) or looping without progress. Argus targets scenarios where the goal itself must co-evolve with the problem being solved—not exam-style questions, but genuine research workflows.

Four-Role Architecture

| Role | Responsibility | |------|---------------| | Manager | Anchors user intent, holds the authority to amend the Work Contract | | Planner | Decomposes the next unit of work | | Engineer | Executes, implements, and evaluates | | Reviewer | Audits artifacts and issues completion verdicts |

Key Design — Work Contract K_t = (ι, o_t, c_t, v_t)

ι (standing user intent) — stable user intent, immutable
o_t (operational objective) — current actionable goal, revised by evidence
c_t (constraints) — known constraints, updated with new discoveries
v_t (verification criteria) — verification standards, refined as understanding deepens

This separation lets the agent revise its operational path based on evidence without tampering with the original intent. The Manager holds amendment authority, but every change must be evidence-backed.

Self-Evolution Through Verification Gating

Model weights are frozen. Evolution occurs in runtime state—memory, skills, verifiers, routing policies—only after passing role review and task-native verification before being "committed."

Candidate generation → accountability check → authorized submission
Failed branches can still become reusable progress
Narrative rewrites that overwrite the entire task without verification are forbidden

Results

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

Token consumption: 1.41x, but mature waves save 21% solve tokens and 15% active time versus startup waves
Recorded 34 verifier recoveries and 22 strict-review rollback rescues
An optimized RWKV6 kernel was merged upstream

Paper Production Pipeline Case Study

Six projects autonomously completed a paper production pipeline:

640 campaign hours
254 bounded tasks
576 Engineer turns
286 Reviewer revisions
16 Stage rollbacks
All 6 pipelines reached submission-complete state

One representative case: 7 no-go rollbacks reframed a failed method search into a 4,500-line negative-results study, then repaired two late-stage submission defects.

One-Line Summary

Argus does not try to make agents smarter—it makes them disciplined: in long-horizon fog, don't alter the goal, don't fake completion, and escalate unfixable problems back to humans.

Reference

https://arxiv.org/abs/2608.05144