Google is opening Antigravity’s /teamwork workflow to external researchers and developers, giving paid-plan users access to a system designed for large software projects, complex simulations and long-running research.
The November 13, 2026 announcement, written by Google Research Fellow and VP Vahab Mirrokni, describes work conducted across Google and academic partner institutions. Google says those collaborations used Gemini models and multi-agent workflows to tackle problems in mathematics, chemistry, systems software and machine learning engineering.
The release is a workflow expansion, not a new Gemini model launch. Google names Gemini 4 Argon, Carbon and Hydrogen among the models used in the research. /teamwork is the coordination system around models: it divides work, assigns specialized roles and checks results across a project.
The Antigravity documentation lists /teamwork as available on paid plans in both the Antigravity App and CLI. Google also says additional capabilities will reach external users in the coming weeks, so the research announcement should not be read as a promise that every specialized research configuration is already exposed to customers.
How /teamwork Organizes Long Projects
Antigravity positions /teamwork for work that can stretch beyond one productive agent session: repository-wide migrations, coupled changes across dozens of files, systems simulations and mathematical research.
Its architecture addresses two problems Google identifies with long-horizon AI work. A model’s first candidate solution can be wrong, and a growing conversation can make important details harder to retain.
Google’s response combines parallel exploration with separate checks on candidate answers. Those checks can include Lean 4, software test suites, execution simulators and hardware counters. Meanwhile, individual agents receive the files and tools needed for their assigned step, keeping their working context narrower.
The documented roles divide responsibility as follows:
- Sentinel records the approved request, routes tasks, provides progress updates and starts the final review.
- Project Orchestrator turns the brief into milestones, coordinates parallel tracks and delegates work. A fresh successor takes over between milestones to limit context degradation.
- Explorers and Workers separate investigation from implementation. Explorers inspect the repository without changing source files; Workers build components, refactor code and write tests in focused, non-overlapping tracks.
- Critic, Challenger and Auditor provide different checks. The Critic reviews correctness and interfaces, the Challenger develops adversarial tests and edge cases, and the Auditor checks integrity requirements and test evidence against actual command output.
A separate Success Auditor performs the final end-to-end verification before Sentinel presents the finished project.
The consequential detail is the separation between producing an answer and accepting it. A Worker’s explanation of why a change works is only one input. Other roles inspect the change, probe failure cases and check execution evidence.
My view is that this separation matters more than the agent count. Adding agents can produce more candidate work; explicit acceptance gates give a project a structure for deciding which work deserves to survive.
Scoping Comes Before Execution
/teamwork begins with a structured interview. The main Antigravity agent clarifies the objective, intended audience, requirements and how success will be checked before the project proceeds.
The documentation’s principle is to specify the desired result without prescribing every implementation step. Each requirement should have an objective check, such as a test suite, reference benchmark, metric script or explicit evaluation rubric.
For a repository migration, a useful brief would identify supported platforms, interfaces that must remain compatible, required tests and acceptable performance limits. That gives the verification roles something concrete to evaluate.
This setup has a practical cost: the user must define success clearly enough for a team to pursue it. I’d accept that overhead for a cross-repository migration. For a small bug with a known reproduction, the same ceremony could consume attention that would be better spent fixing and reviewing the change.
When to Choose /teamwork, /boost or a Subagent
Google describes /boost as a sequential solve-and-verify loop refined from previous trajectories. /teamwork provides a broader project structure, including distributed implementation, specialized review and milestone handoffs.
The documentation also shows that /teamwork adjusts its execution path to the request. Distributed coding is the default for multi-file engineering, simulations and research. A lighter iterative-coding path handles one self-contained change and cannot decompose the task; prompts asking to keep work small or focused can trigger it.
That distinction supports the following practical choices:
| Option | When I’d choose it | Reason |
|---|---|---|
/teamwork | A migration, simulation or research project with dependent milestones | Its documented architecture includes project coordination, separate work tracks and verification gates. |
/boost | A contained coding task that benefits from repeated solving and checking | Google describes a sequential loop, which fits a narrower change without requiring a distributed project. |
| An ordinary subagent | One bounded investigation or implementation task while the main session retains coordination | Delegating a specific job can be sufficient when there is no need for a separate project lifecycle. |
/swarm | Evaluate its specific behavior before choosing it for milestone-dependent work | The cited material does not support a like-for-like account of its scheduling, isolation or review guarantees. Choose /teamwork when its explicitly documented gates are requirements. |
These are workflow-fit judgments based on the published design, not measured speed or reliability rankings. The available evidence does not justify declaring /teamwork universally better than /swarm or a simpler agent.
A useful dividing line is dependency management. If several changes must land together, pass shared tests and satisfy a common acceptance decision, /teamwork has a documented structure for that problem. If the task is to trace one call chain or propose one localized fix, a smaller delegation may be enough.
The execution-path documentation also includes document review, mathematical proof work and an explicitly requested large-team math mode. /teamwork therefore covers more than parallel coding, but its ability to route a small task does not mean every small task needs the full workflow.
What Google’s Research Results Show
Google presents the scientific and engineering results as evidence for its human-AI collaboration approach. They remain first-party claims; the opened evidence does not provide independent replication. The human contribution is central to understanding what those claims mean.
Courtade–Kumar: Researchers Supplied Key Structure
Google says researchers from Google, CUHK, USC and CMU used /teamwork’s Long Proof pattern, implementing Stellar Colosseum in Antigravity, to prove the Courtade–Kumar conjecture in full generality.
The information-theory conjecture dates to 2013. In simplified terms, it concerns compressing a random bit string to one bit when information is observed through a noisy channel. The proposed optimum is a function that simply copies one input bit.
According to Google’s account, Gemini first found counterexamples to a generalized conjecture from earlier CUHK work. The CUHK researchers then proposed a new candidate function, reduced the problem to a 14-variable optimization and supplied a key analytical lower bound.
Gemini 4 Argon and earlier models subsequently partitioned the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates. Google says the full proof was formalized in Lean 4, producing 50 million lines of code.
The sequence matters. Human researchers supplied a crucial reformulation; models then carried out extensive exploration and proof development. Calling this simply an AI solving an open problem would hide the collaboration that made the search tractable.
Frequently Asked Questions
3 questions
1Where Is /teamwork Available?
Google’s documentation lists /teamwork on paid plans in the Antigravity App and Antigravity CLI. The announcement does not mean every specialized research capability is already externally available.
