Agent Development Kit (ADK)

The Agent Development Kit (ADK) is a toolkit made by Google for building and testing AI programs that carry out multi-step tasks on their own, following a person's instructions. Just as a car gets a test drive before it's sent out onto real roads, ADK lets developers run the programs they've built under conditions similar to the real thing, to check that everything works as intended before release.

More recently, it's become possible to test even voice-based conversational programs. An AI playing the role of a fake user speaks in an actual voice, and the program's replies are automatically scored against grading criteria that a person has written in advance. This means developers don't have to listen to every call themselves — they can repeatedly check whether a program still works well each time they change the instructions or switch to a new model version.

In articles, you'll see phrases like "Google's ADK now supports automatic evaluation for live voice agents." Keep in mind that ADK isn't a finished chatbot product — it's a developer tool used to build and test such programs.

To try ADK, you'll need a Python development environment. 1) Install the version that includes evaluation features from the command line. 2) Define your multi-step program in code. 3) Write a file specifying what goals and personality the simulated user should have, and what criteria should be used to score its replies. 4) Specify the voice model to use in a configuration file and run the evaluation — you'll be able to review the results of the actual voice conversation along with its scores.