Google Introduces Gemini Agent Skill, Boosting Coding Success Rate to 96.6%
Google has introduced a so-called “Agent Skill” for the Gemini API designed to close the knowledge gap language models face when SDKs change rapidly. The issue is that AI models, once trained, are unaware of their own updates or the latest best practices. The new skill provides coding agents with up-to-date information on models, SDKs, and example code. In tests across 117 tasks, the success rate of the best model (Gemini 3.1 Pro Preview) increased from 28.2% to 96.6%
Illustration of Gemini Agent Skill improving coding success rate across AI models
Older 2.5 models showed significantly less improvement, which Google attributes to weaker reasoning abilities. However, a Vercel study suggests that direct instructions via AGENTS.md could be even more effective. Google is therefore also exploring alternative approaches, including MCP services.
Success rate of Gemini models with and without Agent Skill across 117 coding tasks: Newer models in the 3.x series benefited more from the skill due to stronger reasoning capabilities compared to older models with weaker reasoning. | Image: Google
Conclusion:
Google’s new Agent Skill for the Gemini API highlights a growing shift toward dynamic AI systems that can access real-time knowledge instead of relying solely on static training data.
Is an AI industry analyst covering major AI platforms, enterprise adoption, and strategic moves by Big Tech companies. His work focuses on how AI systems are deployed at scale and how they reshape products, markets, and user behavior.
ES 
EN