Context · Frontier AI · sweep Tuesday, November 17, 2026
Code agent benchmarks for autonomous software engineering
Evaluation of code agents shifts from saved functional tests to long-horizon tasks, evolving repositories and compliance with repository policies. The bottleneck is no longer the ability to generate correct patches but reproducible verification and governance of autonomous code, which conditions enterprise adoption of software agents.
Title, thesis and gravity assessment written by an AI model, without prior human review.
Gravity30/100 · Background
Why this gravity
Los 11 artículos son preprints de arXiv que proponen nuevos benchmarks (SWE-Bench Pro V2, ExpertCrowd-SWE, SWE-Shift, SWE-CC) para agentes de código; no hay release de modelo, adopción industrial declarada ni métrica de un actor Tier 1 que reordene el sector. La proliferación de benchmarks señala un problema metodológico de verificación, no una discontinuidad material observable hoy.
- Coverage
- 1 feed · 1 independent publisher group (a single publisher group)
- Cross verticals
- Frontier AI, Tech Industry, Startups
- Gravity vs. the previous edition
- −5 (35 → 30)
coverage, not accuracy
Lead source
SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark · arXiv CS.AI / Soham Dan, Jeff Da, Sanyam Bhutani, Miguel Romero Calvo, Ivan Guevara, Fernando Carabedo, Jose Bernal, Julio Reyna, Luis Leandro Monsalve Mendez, Ying Liu, Daniel Yue ZhangTimeline · Madrid time
- 06:00
- 06:00
- 06:00
- 06:00
- 06:00
- 05:00
- 05:00
- 06:00Harvesting SWE Benchmarks from Wild Vibe Coding Sessions · arXiv CS.AI
- 06:00
- 06:00
- 06:00
ShareWhatsApp
Related stories
- 0Persistent memory in vision-language-action models for robotic navigation and manipulation
- 80AI agents discovering and patching critical vulnerabilities in open infrastructure
- 80DeepSeek expansion in agents, APIs and inference infrastructure
- 80Moonshot's Kimi K3.1 Pro and open weights in frontier models
- 72OpenAI offensive cybersecurity models and deployment in government clouds