An unattended 11-hour run of Gemini 4 Argon produced two performance changes that Google accepted into its Agent Development Kit, a Python project spanning more than 200,000 lines, roughly 800 files and 16,892 open-source unit tests.
The agent first read repository-specific guidelines, established benchmarks and traced dependencies across dozens of files. Its accepted changes filtered session events before copying history and delayed loading heavy packages until they were needed. Google measured about a 40% reduction in in-memory session-preparation overhead and a 21% improvement in cold package import time.
Those savings are small beside one model call but compound during long tool-using sessions. In Google’s example, removing repeated copies reduced framework overhead for a 158-turn interaction from 65.6 seconds to 56.1 seconds. The agent ran the complete test suite before finishing.
Human review remained essential. Developers rejected four other modifications because of reliability, compatibility, maintainability or weak measurable benefit. The result is evidence that an agent can navigate and improve an established codebase, not that it can autonomously merge everything it proposes. Review, debugging and validation still determined which changes reached the project.