In an experiment described by Google engineers, an unattended 11-hour run of Gemini 4 Argon targeted the ADK codebase, which spans over 200,000 lines across approximately 800 files and includes nearly 17,000 unit tests. The agent was given a goal to make ADK add no noticeable overhead, preserve the public API, and back each change with a regression test.
According to Google, the agent read repository guidelines, established performance baselines, proposed optimisations, reviewed its own changes, and ran all open-source unit tests before finishing. Two optimisations were accepted: filtering session history events before copying to reduce overhead, and introducing lazy loading to delay importing heavy packages until needed.
Google notes that these tasks are not particularly expensive compared with the overall duration of an agent run, but the improvements compound over long sessions. On a single LLM call the savings are unnoticeable, but across 158 internal turns removing unnecessary session copies cut the full run time from 65.6 seconds to 56.1 seconds.
However, of the modifications proposed by Argon, four were rejected during developer review due to concerns about reliability, maintainability, compatibility, or measurable performance benefits. Google says the experiment demonstrates AI agents' ability to modify existing codebases, but their effectiveness remains limited by the time required for post-coding tasks such as PR review, debugging, and testing.