Google’s Python Agent Development Kit release includes repeated evaluation runs, several ways to compare tool arguments and restrictions on script calls. Developers upgrading applications face changes to accepted inputs and stored sessions. The release also marks components as stable and describes file handling and integration updates for running software agents.
Google’s Agent Development Kit, a framework for building software agents in Python, includes repeated evaluation runs in its 2.14.0 release. Developers can set concurrent runs and combine their results as scores or pass rates. Tool inputs can be checked through exact or case-insensitive matching, numerical tolerance or a language model’s comparison of meaning. Google describes these evaluation controls in its release document.[1]
For scripts executed through an environment, the documented interface supplies structured arguments. The free-form command field is no longer accepted, and scripts must start in their own skill directory using supported file types. Calls retaining the old command field receive an error. Google describes this restriction as addressing a route through which a model could execute arbitrary commands after receiving malicious instructions.[1]
Compatibility notes also cover the values delivered to tools and the storage of sessions. Enumerated tool parameters arrive as members of the declared type rather than raw strings. Existing Firestore sessions remain readable, but sessions written in the release’s format cannot be read by earlier versions after rollback. The document also marks execution environments and established evaluation components as stable.[1]
