Want to read

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

Artificial Intelligence Computer Science

Argus is a persistent agentic runtime designed for long-horizon tasks with sparse feedback. It employs four roles to co-evolve problem statements while keeping model weights fixed. The system uses verification-gated admission to ensure cumulative refinement. Benchmarks show Argus achieves 78% on SWE-Bench Pro, outperforming Direct Copilot. The runtime supports self-evolution, reducing token usage and workflow time in mature waves.

Read the paper

Key takeaways

Argus utilizes a Manager, Planner, Engineer, and Reviewer to execute bounded missions over durable project state.
The system achieves approximately 78% on SWE-Bench Pro compared to 59% for Direct Copilot using 1.41 times the aggregate tokens.
Mature SWE-Bench Waves use 21% fewer solve input tokens and 15% less active workflow time per task than startup Waves.
The runtime recorded 34 verifier recoveries and 22 strict review-loop rescues during observational trajectory analysis.
Argus reaches 76.8% on AARRI-Bench research tasks and demonstrates a 28.0 gap on mathematical data synthesis.
Artifacts include an optimized RWKV6 kernel merged into Flash Linear Attention and six paper pipelines spanning 254 missions.

Reflections

No notes on this one yet. I add reflections as I finish or revisit a paper.