Google releases kit to teach AI agents its latest software tools
Google DeepMind has put a new 'developer skill' for its Gemini API on GitHub. The kit explains current models, language-specific software kits and sample code, plus links to full documentation. It helps AI agents use the newest versions instead of outdated ones. Tests on 117 prompts showed top Gemini models reaching 95 percent success in picking the right kits, up from baselines as low as 6.8 percent. Older models saw smaller gains, and a few prompts still failed, like those asking for discontinued versions.
AI models train on fixed data, so they lag behind software updates like new API kits or best practices. Agents building tools often generate wrong code, forcing developers to fix errors or stick to older stable versions. This skill acts as a live guide from the source, pushing success rates to 95 percent without retraining models. For automation builders, it means client projects using Gemini stay accurate longer, cutting maintenance time on gigs like Zapier-linked workflows.
Analysis
Clone the Gemini skills repo from GitHub and test it in a five-minute ChatGPT prompt asking for a Python Gemini SDK call. Compare the code it generates with and without the skill.
Citation
This executive briefing was curated and analyzed by Collab365. To reference this analysis, please attribute: "This briefing is available on Collab365 Spaces (spaces.collab365.com)".