Gemini 4 Argon cuts ADK session prep time by about 40% in one night

Google did a test of its new Gemini 4 Argon model on a hard task in ADK, its open-source framework for enterprise agents. Argon had to find and repair slow code. ADK has about 205,000 lines of Python code. Argon operated for 11 hours at night with no person. Google reports that the work cut in-memory session prep time by about 40% in long sessions. It also cut the time of cold package imports by 21%.

How it works

Argon got one goal and rule files in the repository folder .agents/skills. The files hold the check rules of the ADK team, for a new engineer. A goal mode kept the goal when the agent had to decrease its memory of the task.

The numbers

Argon made 6 changes across 32 files. The team rejected 4 in a check, and 2 are in production. All 16,892 unit tests passed. A first test with no rule files changed the JSON schema that users see.

Not known

Google made this test on its code with its team, and no other group did it again. The text does not tell which changes Google rejected. It gives no cost for the work.

Insight: Rule files made the difference, not the model

In the first test, Argon had no rule files and broke the conventions of the team. In the next test, with rule files, the work went to production. Copilot and the Starsg API also have the model. Many teams can get the same model, so the written rules of a team read as the part that sets the result. Write your rules as files for agents before you get more model access.