Google has released an “Agent Skill” for the Gemini API that addresses a core limitation of AI coding assistants: language models lack knowledge of their own updates and current best practices after training. The new skill provides coding agents with up-to-date information about current models, SDKs, and sample code, according to a report by The Decoder.
In testing across 117 coding tasks, the top-performing model (Gemini 3.1 Pro Preview) saw its success rate jump from 28.2% to 96.6% when using the Agent Skill. The improvement represents a dramatic enhancement in the model’s ability to generate working code that follows current API conventions and practices.
The performance gains varied significantly across different model generations. Newer models in the 3 series benefited far more from the skill than older models, which Google attributes to their stronger reasoning capabilities. Older 2.5 models saw much smaller improvements, which Google says “comes down to weaker reasoning abilities.”
Skills were first introduced late last year by Anthropic and have been quickly adopted by other AI companies. Google’s implementation specifically targets the knowledge gap that emerges when models are trained on outdated documentation or lack access to the latest SDK updates and coding patterns.
The Agent Skill is now available on GitHub. Google is also exploring other approaches, including MCP services, while a separate Vercel study suggests that giving models direct instructions through AGENTS.md files could be even more effective than the current skill-based approach.
