Google’s new “Agent Skill” for the Gemini API addresses a key limitation in AI coding assistants: the inability to stay updated on their own improvements and industry best practices. This innovative feature provides coding agents with real-time information on current models, SDKs, and example code. In testing, the Gemini 3.1 Pro Preview model achieved a remarkable success rate increase from 28.2% to 96.6% across 117 coding tasks, showcasing the effectiveness of the skill. Unlike older models, which experienced minimal gains due to poorer reasoning capabilities, the newer Gemini models significantly benefit from this update. A Vercel study indicates that direct instructions via AGENTS.md files may enhance performance even further. Google is also investigating additional strategies, including MCP services. The Agent Skill is available for developers on GitHub, marking a pivotal step in optimizing AI-driven coding assistance.
For more insights, subscribe to THE DECODER for ad-free access and exclusive content.
Source link
