Argus: An Evolving AI Runtime

The challenge of long-horizon reasoning in AI agents is being addressed by a novel approach that emphasizes persistence and adaptability. Researchers have introduced Argus, a self-evolving runtime designed to maintain progress when evidence aligns with its strategy and to pivot when faced with failures, hidden constraints, or misaligned objectives.

Persistent State and Role-Based Execution

Argus operates with a distinct architecture that separates stable user intent from dynamic operational goals, constraints, and verification criteria. This separation is managed by distinct roles: Manager, Planner, Engineer, and Reviewer, each executing bounded missions over a durable project state. Crucially, new elements like memories, skills, procedures, verifiers, and routing decisions are admitted only after role-specific review and, where possible, task-native verification. This structured approach ensures that the evolution of the agent's capabilities is controlled and validated.

Self-Evolution Without Retraining

A key innovation of the Argus AI runtime is its ability to self-evolve while keeping model weights fixed. This evolution occurs through the persistent runtime state and the control policy, allowing for autonomous execution between operator-defined escalation points. This contrasts with traditional methods that require extensive retraining to adapt to new information or tasks. The system demonstrated this capability across seven GPT-5.5 benchmark arenas, achieving approximately 78% on SWE-Bench Pro, a significant improvement over the 59% achieved by Direct Copilot, albeit with 1.41 times the aggregate tokens.

Quantifiable Improvements and Real-World Applications

The impact of Argus's self-evolutionary process is evident in its performance metrics. After verification-gated self-evolution, mature SWE-Bench waves required 21% fewer solve-input tokens and 15% less active workflow time per task compared to initial waves. These mature waves also recorded 34 verifier recoveries and 22 strict review-loop rescues, highlighting the system's resilience and ability to correct its course. Further validation came from achieving 76.8% on AARRI-Bench and a 28.0-point lead in mathematical data synthesis. Beyond benchmarks, Argus has been applied to practical tasks, including merging an optimized RWKV6 kernel upstream and managing complex multi-day mathematics campaigns. Its success in completing 254 missions across six paper pipelines, with 25 stage rollbacks, underscores the potential of a fixed-weight, self-evolving harness to refine, recover, and accumulate verified approaches, paving the way for future supervised and reinforcement learning advancements.