Google's Gemini 4 Argon optimizes its Agent Development Kit Python codebase overnight

Google says an 11-hour unattended run of Gemini 4 Argon on its Agent Development Kit (ADK) produced two accepted optimizations. The codebase spans more than 200,000 lines across roughly 800 files and includes nearly 17,000 unit tests. The engineers wanted to see how the model handles existing code rather than greenfield projects.

Before the long run, the team wrote repository guidelines in .agents/skills files, which the agent loaded only when a task needed them to keep its context window small. The goal was set with Antigravity's /goal mode so it survived context compaction. The prompt asked for one pull request against main with before and after numbers, and told the agent not to merge it.

The accepted changes filter session events before copying them and lazily load heavy packages. Google reports about 40% less in-memory session preparation overhead in long sessions and 21% faster cold package imports. On a 158-turn tool-calling run, time fell from 65.6 seconds to 56.1 seconds. Still, four of Argon's proposed changes were rejected by developers over reliability, maintainability, compatibility or unmeasured benefit.

Agent developers get a concrete pattern: put repository conventions in .agents/skills files and give the agent a goal with hard constraints, such as an unchanged public API and a regression test for each change. Human review remains the bottleneck, since four of the model's changes were turned down, so budget review time before trusting agent-produced pull requests. Teams building on ADK can expect lower framework overhead in multi-turn tool calling.