A study of 34,000 AI agent skills reveals that modular instructions designed to enhance performance barely help in realistic conditions. Weaker models often perform worse when using these skills than without them.
AI agents are built to access specialized knowledge through skills—modular instructions that can be deployed dynamically to improve performance. However, researchers testing 34,000 real-world skills found the enhancement strategy largely ineffective outside controlled benchmarks.
The gap between benchmark performance and real-world results suggests current skill implementations don't translate well to practical scenarios. The findings raise questions about how agent architectures handle skill integration and deployment.
Weaker models showed particularly poor results, performing worse with skills enabled than without them. This indicates skills may introduce complexity that smaller models struggle to manage effectively.
The research highlights a common challenge in AI development: techniques that show promise in standardized tests often underperform in production environments. As AI agents move toward broader deployment, bridging this performance gap will be critical for practical applications.
India's JioStar is integrating generative AI into its streaming platform to enable conversational recommendations for shopping and entertainment. The move positions AI-powered interactions as a core revenue and engagement driver.
Cognition has released SWE-1.7, a new AI model trained using Kimi K2.7 that processes text at 1,000 tokens per second. The company claims the model matches performance of GPT-5.5 and Claude Opus 4.8 while reducing costs.
Westpac Banking Corp. is ramping up oversight of artificial intelligence costs by tracking token usage across the organization and directing routine tasks to cheaper models.
Google's Gemini app can now generate lifelike videos featuring AI avatars of users. The technology creates digital clones that mimic appearance and behavior with striking accuracy.