Google says its Gemini 4 Argon model raised agentic code submissions at its own teams by more than 35%, but the figures are internal and unverified. Separately, Microsoft researchers describe AI performance degrading in longer, multi-turn work, and Linux kernel reporting suggests AI use in code is often not disclosed.
How this is computed
- Key claims supported: 10 of 24
- Type of source: –
- Independent confirmation: None
Rates the source, not the topic and not our analysis.
Why it matters
Our reading: the sources do not address each other, but together they illustrate a gap between benchmark and vendor-reported gains and what can be measured in everyday software work. Without independent data on Google’s internal gains, and with AI use in a major open-source project largely untagged, the productivity effect of coding agents remains hard to assess from outside.
What the sources say
- Vendor claim only Google DeepMind says Argon in its Antigravity tool helped Google teams increase agentic code submissions by more than 35% and cut code rollbacks by nearly a third. No method, baseline, time frame or sample is given.
- Vendor claim only Google DeepMind reports a score of 77.9% on the DeepSWE v1.1 benchmark for long-horizon software engineering and calls it a new state of the art, without linking methodology or independent results.
- Conditionally supported Jennifer Neville of Microsoft Research says models’ performance on benchmark tasks re-specified over several turns with simulated users “actually degrades significantly over multiple turns”. The underlying paper’s method and review status are not stated in the interview.
- Conditionally supported Neville also says errors in long agentic workflows, such as repeated document edits, can accumulate and cause surprising failures. This is the kind of long-horizon work Google’s benchmark claims target, though she does not comment on Argon.
- Conditionally supported Kernel developer Sasha Levin says only about 10 percent of the 18,533 commits in Linux 7.3 through -rc7 carry the AI-assistance tag. He adds that AI is used much more often, an estimate without published data.
- Supported heise online reports that Linus Torvalds stated in a commit message that code, test programs and a first text draft were created with Claude Code under his guidance. The claimed speedup of that code is not benchmarked in the article.
The key caveat
The most concrete productivity figures come from Google about its own teams, with no method or baseline published; none of the sources offers an independent measurement of coding-agent productivity.
Open questions
- What baseline, time frame and team sample underlie Google’s reported 35% rise in agentic code submissions?
- Do the multi-turn degradation findings from Microsoft Research apply to Argon or other current coding agents?
- If AI disclosure moves out of commit messages, how could AI’s share of real-world code be measured?
What’s next
Watch whether the patch removing the kernel’s AI-tag requirement, already accepted into the docs-next branch, reaches the mainline kernel.
Sources: Google DeepMind Blog, 30 September 2026: Gemini 4 Argon: our next era of frontier intelligence · Microsoft Research Blog, 6 October 2026: What AI gets wrong and what failure teaches us · Google DeepMind Blog, 13 October 2026: Gemini 4 Argon: Built to solve your hardest problems · heise online, 16 October 2026: Linux-Kernel: Maintainer wollen „Assisted-by“-Tag streichen
Produced by Zeitgeist’s AI agents (research, editing, standards check) with Claude. No human review.