Google introduces two tools designed so code agents stop producing outdated solutions and start always using the correct version of APIs and SDKs.
Qué problema resuelven
Have you ever had an agent return code that no longer compiles or uses outdated parameters? That happens because many agents were trained on data with a cutoff date. The result: examples, function signatures, and configurations that no longer match current versions.
This isn’t a small glitch. A renamed endpoint, a removed parameter, or a change in best practices can turn a useful answer into a headache for any developer or team. Think of getting a recipe that calls for an oven setting your model doesn’t even have anymore — frustrating, right?
Las dos herramientas y cómo funcionan
Gemini API Docs MCP
Gemini API Docs MCP connects your code agent to live Gemini documentation: APIs, SDKs, and model details. It works via the Model Context Protocol to inject up-to-date context into the agent just before it generates code.
The result? Your agent no longer relies only on its prior training; it can consult the most recent source. If a function signature changed or a new recommended configuration appears, the agent will know.
You can see the gateway at https://gemini-api-docs-mcp.dev.
Gemini API Developer Skills
This second piece adds best-practice instructions, useful links, and recommended usage patterns to guide the agent toward producing code that matches current SDKs. It doesn’t just inform — it teaches the agent how to structure solutions according to prevailing patterns.
Think of it as a style guide and set of practical recipes that prevent an agent from applying outdated habits even when it has access to the docs. For example, it can prefer a newer auth flow or the recommended client initialization for a given SDK.
Por qué funcionan mejor juntas
Each tool improves the workflow on its own: MCP brings the technical truth, and Skills teach the right way to use it. But combined they reach their full potential.
In Google’s tests, using MCP and Skills together hit 96.3% success on their evaluation set. Also, correct responses used 63% fewer tokens on average compared to traditional prompting techniques. Fewer tokens means more concise answers and potentially lower inference cost — which matters if you run many queries.
Cómo aprovecharlas hoy
- Connect your agent to the Gemini API Docs MCP to give it dynamic access to documentation.
- Enable the Gemini API Developer Skills to enforce patterns and good practices in code generation.
- Evaluate with your test set and tweak prompts if you need to prioritize style or performance.
If you want to implement it now, Google provides a setup guide at ai.google.dev/gemini-api/docs/coding-agents.
Reflexión final
The lesson is clear: agents shouldn’t be black boxes that recite what they knew during training. If you give them access to current documentation and a guide of best practices, they produce more accurate, efficient, and less costly solutions. Isn’t that exactly what we want when we automate parts of development?
Fuente original
https://blog.google/innovation-and-ai/technology/developers-tools/gemini-api-docsmcp-agent-skills