✓ ShippedByAgents
Vendor-published Refactor / maintenance

Gemini 4 Argon runs unattended for 11 hours on the 205,000-line ADK Python codebase; 2 of its 6 optimizations ship

For the Gemini 4 Argon launch, Google engineers had the model work overnight in Google Antigravity on google/adk-python (about 205,000 lines, nearly 17,000 unit tests) with a one-paragraph goal: cut ADK framework overhead without changing the public API. A first two-hour run found real bottlenecks but broke maintainer conventions. With maintainer guidelines added as .agents/skills and /goal mode on, the second run lasted 11 hours, ran all 16,892 unit tests and opened one pull request with six optimizations across 32 files. Reviewers shipped two and rejected four.

The numbers (as reported)

Codebase~205,000 lines, ~800 source files, ~17,000 unit tests
Run length11 hours unattended (after a 2-hour first round)
Proposals6 optimizations across 32 files; 2 shipped, 4 rejected in review
Session prep overheadAbout 40% lower in long sessions
Package import time21% faster

Facts

Company / subject
Google
Tools
Google Antigravity
Models
Gemini 4 Argon
Codebase
brownfield
Work type
Refactor / maintenance
Stack
Python
Published
2026-10-13
Last checked
2026-10-21

What the source claims Vendor-published

Claimed by the vendor that sells the tool. Not checked independently.

  • Session prep time cut by about 40% in long sessions and cold package imports 21% faster, after an 11-hour unattended run -40% prep; 21% faster imports
    “cut ADK's in-memory session prep time by about 40% in long sessions and sped up cold package imports by 21%”
  • A 158-turn run went from 65.6 to 56.1 seconds 65.6 s → 56.1 s
    “cuts the full 158-turn run from 65.6 seconds to 56.1 seconds”

Caveats

Vendor launch post for its own model. Time to first model request in a new process stayed flat (3.07 s to 3.09 s). No cost figures.

Related topics: Refactoring and maintenanceBrownfield codebasesVendor case studies

Source

Google Developers Blog

One email a week: what actually shipped

New sourced stories, the numbers behind them, and what didn't work. No hype.

We'll only use this to send the newsletter. See the privacy page.