Google says Gemini 4 Argon-assisted changes cut ADK overhead

On October 13, Google published an account of Gemini 4 Argon working on its open-source Agent Development Kit (ADK). Google says two changes proposed during an 11-hour run passed human review and reduced selected ADK framework overhead. The results measure time spent preparing work inside ADK—not Argon’s general coding ability or the model’s response speed. (Google’s ADK account)

Google tested the changes in google/adk-python, a repository of roughly 205,000 lines and nearly 800 source files. NeoTeo previously covered reported questions around Argon’s practical coding performance; the ADK work offers a narrower technical case.

What Google’s ADK test measured

Framework overhead is the time ADK spends setting up and handling a request around the model call. To isolate that work, Google says it used instant model and tool responses, identical model requests before and after the code changes, and medians from 6–12 runs. The holdout benchmark was created before the runs and not shown to Argon.

The long-session numbers measure ADK’s preparation time before the model responds. Across the full 158-turn run, Google reports a 9.5-second reduction from the 65.6-second baseline, or about 14.5%. The cold import also became faster, but the time from starting a new process to its first model request stayed effectively flat: Google says the deferred modules load on that first turn.

Which ADK paths changed

Google says Argon proposed six changes and human reviewers accepted two. One filters in-memory session events before copying the history; the other defers selected heavy package imports until they are needed while keeping public attributes patchable. Together, the changes reduced long-session preparation overhead by about 40% in the InMemorySessionService path.

That result applies to in-memory sessions. Google says the change did not improve session-store read times for external databases, SQLite, the Gemini platform, Firestore, or Redis.

The repository work involved more than producing a patch. Google says its first, roughly two-hour attempt lacked custom repository guidance and missed maintainer conventions, including a public JSON schema and mock-event edge cases. For the second attempt, Google added guidance files and used /goal mode in Google Antigravity. Argon read 12 guidance files during the 11-hour run; Google says it ultimately reran all 16,892 open-source tests. Reviewers accepted two of the six proposed changes and rejected four.

Argon’s coding results vary by benchmark

Google reports a 77.9% score for Argon on DeepSWE v1.1. (Google’s announcement) But other coding suites yield a more mixed picture. MindStudio reports 55% on Frontier SWE v2 for Argon, compared with 65.5% for GPT-6 Astra and 62.3% for Claude Opus 5.5. In MindStudio’s Terminal-Bench comparison, Argon scores 57.4% and Claude Opus 5.5 scores 66.4%. MindStudio notes that the DeepSWE comparison draws on different evaluation setups: Argon’s result used Google’s agent harness, while competitor figures came from separate leaderboards and system cards. (MindStudio’s benchmark analysis)

Vals AI reports 68.17% ± 4.35 for Argon on Code Migration and 91.91% ± 1.90 on Vibe Code Bench v1.1. It lists Claude Sonnet 5.5 at 69.83% on Code Migration and 92.39% on Vibe Code Bench v1.1. (Vals AI’s results) Each suite measures a different set of coding tasks, so these scores do not produce a single overall winner. Argon’s reported DeepSWE result is strong, but it does not establish consistent superiority across software engineering, terminal work, migration and app-building tasks.

What Navin Reddy’s project adds

Telusko creator Navin Reddy says he received early access to Argon and used it while building and iterating on CamStudio Recorder, an application of roughly 20,000 lines written in Rust, React and WebGL. He also describes using the model on audio/video synchronization problems. That is one developer’s account of work on a substantial project, separate from Google’s ADK measurements.