AI Research

Argus: A General-Purpose Agentic Runtime for Long-H... | AI Research

Key Takeaways

Argus is a general-purpose agentic runtime designed to handle long-horizon research tasks that require persistent reasoning, evidence-backed pivoting, and se...
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.
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.

Paper AbstractExpand

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.

Argus is a general-purpose agentic runtime designed to handle long-horizon research tasks that require persistent reasoning, evidence-backed pivoting, and self-evolution. Unlike agents that operate under a fixed objective, Argus allows for the refinement of operational goals and constraints as a project progresses, provided these changes are supported by evidence and pass through a structured review process.

The Argus Runtime Architecture

The system organizes work into bounded missions executed against a durable project state. Four distinct roles manage the workflow:

Manager: Anchors the standing user intent and oversees campaign state and stage transitions.
Planner: Determines the next units of work.
Engineer: Implements tasks and performs evaluations.
Reviewer: Inspects artifacts and outputs to issue completion verdicts.
This structure separates the stable user intent from the operational objective, constraints, and verification criteria. By using a "verification-gated" approach, the system ensures that memories, skills, procedures, and routing decisions are only added to the persistent state after they have been reviewed and verified. This allows the system to "self-evolve" its control policy even though the underlying model weights remain fixed.

Performance and Benchmarking

Argus was evaluated across seven benchmark arenas, including software engineering, GPU kernel optimization, and mathematical data synthesis. On SWE-Bench Pro, Argus achieved approximately 78% accuracy compared to 59% for Direct Copilot, while utilizing 1.41 times the aggregate tokens.
The researchers observed that as the runtime self-evolved through verification-gated cycles, mature stages of the SWE-Bench tasks became more efficient, using 21% fewer solve-input tokens and 15% less active workflow time per task than startup stages. The system also recorded 34 verifier recoveries and 22 review-loop rescues during these runs.

Research and Practical Applications

Beyond standard benchmarks, Argus has been applied to multi-day research campaigns. In one instance, an optimized RWKV6 kernel produced by the system was merged into the upstream Flash Linear Attention repository. In a separate mathematical campaign, the system retained a falsified route and six proof-backed frontier updates. Six paper-production pipelines completed 254 missions, utilizing 16 stage rollbacks to reach submission completion without resetting the research state.

Considerations for Long-Horizon Reasoning

The authors note that long-horizon research often involves objectives that are underspecified or misspecified at the start. Argus addresses this by permitting "verified pivoting," where the system can revise its operational objective based on evidence. This is intended to prevent "goal drift," where an agent might otherwise rationalize failure by degrading its objective to match its current capabilities.
The system relies on the assumption that verification is not merely a final quality filter but a mechanism to distinguish rationalized failure from legitimate objective refinement. The retained trajectories from these missions are intended to serve as structured data for future supervised and reinforcement learning.

Comments (0)

No comments yet

Be the first to share your thoughts!