Google Research reported new results from Antigravity’s /teamwork system on November 13, 2026, describing collaborations between Gemini agents and human experts across mathematics, science and software engineering. The announcement gives university research teams examples of extended agent workflows to examine, while distinguishing tools available externally from specialized research patterns still due to arrive.
What Google announced
In the Google Research report, Google Fellow and vice president Vahab Mirrokni describes campaigns using Gemini 4 Argon, Carbon and Hydrogen with multi-agent workflows. The company says human researchers frame problems and constraints while agents explore candidates, build computational pipelines and check results.
The November 13 development is a report of recent work and external access, rather than the first introduction of Teamwork. Google explicitly builds on earlier posts. It also separates general availability from forthcoming specialized workflows, an important distinction for researchers seeking to reproduce a particular demonstration.
Engineering results remain Google’s claims
One concrete example is the migration of the Zircon microkernel from C++ to Rust. Google reports that a two-week campaign completed the remaining code after an initial quarter had been ported, adding about 800,000 lines of Rust. Its stated acceptance evidence is that the result passes kernel unit tests and boots on arm64 hardware.
In MACCUS’s November 15 coverage, Marc Winer distinguishes those checks from production deployment. His examination notes that Google’s announcement does not identify whether the kernel runs in a product or has completed the audits discussed in its earlier reporting.
Winer also examines Google’s libgav1 video-decoder rewrite. Google reports output matching across a conformance suite and a final Rust implementation running 20–30% faster on sample clips after profiling. MACCUS emphasizes the restricted comparison and the absence of published new decoder code in the announcement. These figures describe selected company campaigns; they do not establish a general performance gain for research software rewritten with agents.
How the public workflow is organized
The public Teamwork documentation describes the command as available through paid Antigravity app and CLI plans. Projects begin with a scoping interview that establishes objectives, requirements, acceptance criteria and an independent check for each requirement.
Execution then proceeds through milestones. Google documents separate implementation and review roles: critics inspect correctness and interfaces, challengers construct adverse tests, and auditors check evidence against actual command output. Project artifacts record the request, plan and progress. These are product descriptions, rather than independently measured guarantees of error-free work.
The documentation also distinguishes three integrity modes. Development permits code reuse and libraries; demo adds restrictions; benchmark requires an independent implementation using the language’s standard library. The selected mode follows the shortcuts ruled out during scoping. A university comparing runs therefore needs more than the product name to understand the evaluation conditions.
Release notes provide separate context
Google’s Antigravity changelog lists app version 2.32.0 on November 13, including Dynamic Workflows for orchestrating agents and multi-step work through a bundled Python runtime. The changelog warns that new versions roll out gradually and may take several days to reach all users.
It separately records Gemini 4 Carbon on November 10 and Hydrogen on November 12. Those entries concern model access and product changes. They should not be treated as independent verification of the scientific and engineering results described in the November 13 research report, or as a promise that every research workflow is already exposed to every account.
Independent analysis stresses evaluation boundaries
In its November 15 analysis, the Learnetto team argues that a useful external checker matters more than the number of agents. Its account distinguishes generating plausible candidates from rejecting incorrect ones through tests, proof checking or measured execution.
Learnetto also identifies missing denominators: Google’s announcement does not supply an aggregate count of attempted problems, unsuccessful campaigns or total resource consumption. The outlet cautions that the configuration available to an outside user may differ from the internal configuration behind a selected result. Both Learnetto and MACCUS offer analysis of Google’s disclosure; neither reports independently reproducing the campaigns.
What this means for university research teams
For research offices, our assessment is that three separate questions now matter: whether a researcher can access the service, whether the required workflow is available, and what evidence would establish success on the institution’s own task. The announcement answers parts of the first two; a laboratory’s evaluation must answer the third.
An AI evaluation card provides a place to record model configuration, acceptance criteria and the scope of testing. For a software migration, the difference between passing a specified suite and demonstrating equivalent behavior outside that suite belongs in the research record. That distinction connects these reports to computational reproducibility.
The public workflow’s review roles also make research software code review a concrete part of the adoption question. Our interpretation is that institutions should assess the evidence produced by those roles alongside human review, rather than infer equivalent controls across runs using different integrity modes.
What happens next
Google says applications are open for 10,000 academic researchers to receive six months of access to its highest usage tier. It describes the formal-proof pattern and additional specialized workflows as forthcoming, with tutorials planned in the coming weeks. The November 13 post gives no specific release date for those additions. The immediate development is the results disclosure and access opportunity; further workflow releases remain planned.
Sources
- Google Research: Advancing science and engineering with Antigravity’s /teamwork, Vahab Mirrokni, November 13, 2026. Checked 2026-11-15.
- Google Antigravity Docs: Teamwork agent teams, undated documentation. Checked 2026-11-15.
- Google Antigravity Docs: Changelog, entries dated November 10–13, 2026. Checked 2026-11-15.
- MACCUS: Google’s AI agents finished porting an OS kernel to Rust in two weeks, and it boots, Marc Winer, November 15, 2026. Checked 2026-11-15.
- Learnetto: Google reports research gains from Gemini agent teams, Learnetto team, November 15, 2026. Checked 2026-11-15.
