Gemini 4 Argon's 11-hour unattended run yields two accepted optimizations in Google's ADK codebase

Google says an 11-hour unattended run of Gemini 4 Argon produced two accepted optimizations for its Agent Development Kit (ADK). They reduced in-memory session-preparation overhead by about 40% in long sessions and cold package import time by 21%. InfoQ reported the details on October 24, 2026.

The test: an existing codebase, not a greenfield app

Google engineers say they wanted to learn "about how the model handles an existing codebase", which they describe as a much harder task than building "greenfield apps built from an empty folder". They pointed the model at ADK itself: more than 200k lines of code across approximately 800 files, with nearly 17,000 unit tests.

Performance work was chosen because the slowdown rarely sits where the symptom appears. The model has to hold a map of the whole codebase over a long run.

Setup

An initial short run showed the team what the prompt needed to include so the agent would follow its unit-testing and benchmarking conventions. Engineers then wrote repository-specific guidelines as .agents/skills files, which Google compares to an onboarding guide for a new engineer. The agent was told to load only the skill required for the task, to avoid flooding the context window.

The final prompt used Antigravity's /goal mode so the objective would survive context compaction. It read:

/goal Make ADK add no noticeable overhead to any agent: at startup, in long sessions, in multi-agent and workflow setups. Keep what the model receives and the public API unchanged, and back each change with a test that would catch a regression. Open one pull request against main that the ADK maintainers would approve, with before and after numbers in its description. Don't merge it.

Results

Google says Argon's logs show the agent read the repository guidelines, established performance baselines, proposed optimizations, reviewed its own changes, and ran all 16,892 open-source unit tests before finishing.

Two optimizations were accepted into the codebase:

• Session history: events are filtered before they are copied. Google reports this made in-memory session preparation about 40% faster in long sessions.
• Lazy loading: heavy packages are no longer imported until needed. Google reports this cut cold package import time by 21%.

Google argues the agent did cross-module work a profiler cannot do. A profiler flags that copying sessions and importing packages take time. Argon, according to Google, traced every caller, mock target and shared reference across 40 files so the faster paths would not break existing apps.

Why small gains matter in agent loops

Google notes that ADK supports multi-tool interactions spanning dozens or even hundreds of iterations. In its words, a human won't feel 28 ms or 78 ms on a single LLM call that takes seconds. Across 158 internal turns of tool calling, however, the copies compound. Removing them cuts the full 158-turn run from 65.6 seconds to 56.1 seconds, which Google describes as nearly 10 seconds of pure framework overhead.

Rejected changes

The run was not clean. Argon also produced 4 modifications that were rejected on developer review, over concerns about reliability, maintainability, compatibility, or measurable performance benefits. InfoQ notes that the effectiveness of AI agents on existing codebases is still limited by the time and effort of post-coding work such as PR review, debugging and testing.

Gemini 4 Argon's context window, pricing, parameter count and benchmark scores were not stated in the report.

What this means

The account shows an agent working on a large, mature Python codebase with a test suite as its safety net. Repository-specific skill files, a goal that persisted across context compaction, and a requirement to back each change with a regression test shaped the result. The prompt also required before-and-after numbers in the pull request, which gave reviewers measurements to check.

The outcome also shows where the cost of this workflow lands. Two changes were accepted and 4 were rejected, so human review still decided what shipped. The results are Google's own, on its own codebase, and come from a single reported run.