Google Developers unveils a guide on creating self-learning AI agents using ADK SkillToolset

The Google Developers team has presented a guide on creating agents using the Agent Development Kit (ADK) SkillToolset. The architecture is based on various skill interaction patterns, including the use of file-based and external modules. Developers can locally load configuration files via specialized directories, making the knowledge base reusable. Additionally, it supports importing external solutions from custom repositories, as well as official Google skills that are installed via console commands. The format of all these modules strictly adheres to the open specification agentskills.io, ensuring their universality and basic compatibility.

The most advanced level of integration in ADK is the concept of meta-skills, which allows an agent to expand its capabilities without direct human intervention. Using built-in instructions and access to the specification via the resources field, the agent can autonomously generate new valid configuration files. This system functions as a kind of "Skill Factory": upon receiving a request to create a tool for Python code security analysis, the agent studies reference examples and forms a ready-made module. A skill generated in this manner will include structured instructions covering input validation, authentication, and cryptography.

Format unification plays a key role in scaling this technology beyond a single ecosystem. Since modules created by an agent or written manually conform to a single standard, they work not only within ADK. The agentskills.io specification is already supported by over forty different products in the artificial intelligence market, including solutions like Gemini CLI, Claude Code, and Cursor. The source materials do not provide a complete list of all compatible platforms, but it is emphasized that the standardized approach allows for free exchange of components between different development environments.

For the final assembly of a working solution, all created or imported skills are combined into a single SkillToolset package, which is then passed to the root agent based on the gemini-2.5 — flash language model. Despite the high level of automation and the ability of AI systems to self-expand, Google specialists strongly recommend maintaining human involvement in the development process. Engineers are advised to personally review generated files and conduct thorough testing of their effectiveness before implementation. Such control ensures that skills created by algorithms will operate precisely according to specified requirements.