Google Research has published results from Antigravity’s /teamwork, a multiagent system that pairs Gemini 4 models with human experts on open problems in mathematics, chemistry, systems software, and ML engineering. The write-up reports work in information theory, formal verification, molecular-structure discovery, kernel optimization, and production infrastructure. Google says the same tooling is now reaching outside users, alongside /boost and other agents.
That combination is the story. Plenty of labs show an AI model solving a puzzle. Far fewer describe a team of agents, checked by proof assistants and test suites, working next to people on problems nobody has closed yet. In the material reviewed for this report, Google Research does not attach benchmark scores, speedup factors, or named researchers to those results, so this analysis separates what the announcement states from what it leaves open.
The timing matters too. Gemini 4 Carbon, Gemini 4 Argon, and Gemini 4 Hydrogen all reached developers this month, and our earlier look at Carbon’s Lean-verified code work showed how central formal checking has become to Google’s pitch. /teamwork is the layer that turns three separate models into one coordinated crew.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What Google Research Reported on Antigravity’s /teamwork #
The Google Research post frames /teamwork as a multiagent system built to support human-AI collaboration. It does not present the agents as replacing researchers. The work is described as done in collaboration with human experts, which shapes how readers should interpret every result in it.
Four fields get named: mathematics, chemistry, systems software, and ML engineering. Those are very different disciplines, and they reward different kinds of correctness. A proof either checks or it does not. A molecular structure has to match physical data. A kernel has to run faster on real hardware without breaking behavior. A production change has to survive live traffic.
The reported application areas map onto those fields. Google lists information theory and formal verification on the math side, molecular-structure discovery on the chemistry side, kernel optimization for performance work, and production-infrastructure work inside real systems. Google also says the models and tools are being made available externally through /teamwork, which means developers and scientists can try the approach instead of reading about it.
One caution applies from the start. The announcement describes areas of work, not a leaderboard. Readers hoping for a single number that proves the system works will not find one in the reviewed material.
How /teamwork Differs From /boost and Single Agents #
Antigravity’s documentation draws a clear line between its agent modes. The /boost command activates an on-demand multi-agent reasoning pipeline for hard software-engineering tasks, such as stubborn bugs, algorithm problems, race conditions, and tricky refactors. The /teamwork command targets something bigger: large software projects, complex simulations, multi-milestone work, repository-scale migrations, and deep research.
The documentation also describes how a /teamwork session starts. It opens with a scoping interview, and only then does the system build a project plan. A coordination layer built around a Sentinel and a Project Orchestrator manages specialized roles and review gates. That structure resembles a small engineering department more than a chatbot, and it explains why Google can aim the tool at open problems instead of single prompts.
For a working engineer, the practical difference is scope and duration. /boost fits an afternoon of debugging. /teamwork fits a project that spans days, with milestones and checkpoints. The Antigravity docs list subagents and a swarm mode as further options, so the product now offers a ladder from one delegated side task up to a full coordinated team.
Google Research ties the science results to that second tier. The research blog describes multiagent systems running inside /teamwork, not a lone model answering a question, and the distinction is the whole point of the release.
The Three Gemini 4 Models Behind the Work #
Google names Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen in the announcement. Each plays a different part, and Google’s model documentation spells out the split. The table below summarizes how Google describes them, drawing on its AI Studio and model-card pages.
| Attribute | Gemini 4 Carbon | Gemini 4 Argon | Gemini 4 Hydrogen |
|---|---|---|---|
| Google’s positioning | Strongest reasoning model | Efficient frontier reasoning model | Fast execution for everyday agentic work |
| Stated strengths | Sustained agentic workflows, scientific reasoning, complex software engineering, cybersecurity | Lower token use and faster responses than Carbon on coding, agentic, math, and science tasks | Rapid execution, can act as worker or verifier next to Carbon |
| Availability (November 2026) | AI Studio, API, Gemini for business, Antigravity, Gemini app, AI Mode in Search | Gemini API and AI Studio | Introduced in November 2026 |
| Listed API price per 1M tokens | Not covered here | $2 input and $10 output (intro through November 30), then $4 and $20 | $1.50 input and $7.50 output |
| Context window | Not covered here | Not covered here | 1 million tokens |
| Likely role in a team | Hard reasoning and planning | Deep reasoning at lower cost | High-volume workers and checkers |
Carbon as the heavy thinker
Google calls Carbon its strongest reasoning model, built for long-horizon agentic work across software engineering, multimodal knowledge work, scientific reasoning, and cybersecurity defense. Availability varies by product and account tier. A multiagent team with open scientific problems on its desk would naturally put Carbon on the hardest reasoning steps.
Argon as the efficient reasoner
Argon aims for frontier intelligence with fewer tokens per turn. The introductory price runs through November 30, and the listed rates double after that date. For a long /teamwork run, token spend adds up quickly, so that price step is worth planning around. Our report on Gemini 4 Carbon’s $2 intro price and uneven rollout covers a related availability wrinkle.
Hydrogen as the worker and verifier
Google describes Hydrogen as built for speed, a model that can stand alone or pair with Carbon and Argon as the high-efficiency execution engine. That fits a team design where many cheap workers run in parallel while a stronger model plans and judges. We covered the launch pricing in our piece on Gemini 4 Hydrogen’s $1.50 input price and 1M context.
Four Domains: Mathematics, Chemistry, Systems Software, ML Engineering #
Google Research groups its results by field, and each field tests a different strength of a multiagent team. Mathematics rewards precision. The announcement points to information theory and formal verification, two areas where a claimed result can be checked line by line by a machine.
Chemistry brings messier targets. The announcement reports molecular-structure discovery, a task where candidate answers must agree with experimental evidence. Agents can propose and filter structures at a pace no human group could match, though a chemist still has to decide which candidates deserve lab time.
Systems software and ML engineering share a feature: measurable output. Kernel optimization shows up in the announcement, and a kernel either runs faster on the target chip or it does not. Production-infrastructure work gets a mention as well. Google ties the team approach to real operational systems, not only to research demos.
The spread across four fields is deliberate. A system that only handled code could be dismissed as a better autocomplete. A system that touches proofs, molecules, and kernels makes a broader claim about general problem-solving, and that claim now needs outside testers to confirm or deny it.
Why Lean 4 and Test Suites Matter for Trust #
The most interesting detail in the announcement is not a model name. It is the list of verification tools: the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, and hardware counters. Agents that grade their own work are weak evidence. Agents whose output must pass an external checker are much stronger evidence.
The word unmodified deserves attention. If an agent can edit the tests it must pass, a green result means little. Running the original suite closes that loophole. The table below is our own analysis of what each named check can and cannot establish. Google does not publish this mapping.
| Named check | What it can confirm | What it cannot confirm |
|---|---|---|
| Lean 4 proof checker | A formal proof follows from its stated axioms | That the theorem statement captures the real-world question |
| Unmodified software test suites | Existing behavior still passes after a change | Behavior the suite never tested |
| Execution simulators | A design behaves as expected in a modeled environment | Fidelity of the model itself |
| Web search | A claim matches published literature | That the literature is correct |
| Hardware counters | A kernel’s measured performance on real silicon | Performance on other chips or workloads |
| Human experts | Relevance, taste, and scientific judgment | Exhaustive coverage at machine speed |
The takeaway: every tool in the list narrows the room for confident nonsense, but none closes it. That is why the human-expert role in the announcement is more than a courtesy.
What the Announcement Does Not Say #
A fair read of this story requires a clear list of gaps. In the source material reviewed here, Google Research gives no specific performance metrics, no prices for /teamwork itself, no named human speakers, and no confirmed launch date for the research post. It offers no on-record quotations that we could verify.
| Item | Status in reviewed sources |
|---|---|
| Models named: Argon, Carbon, Hydrogen | Confirmed |
| Fields: math, chemistry, systems software, ML engineering | Confirmed |
| Verification tools incl. Lean 4 and unmodified test suites | Confirmed |
External availability via /teamwork alongside/boost |
Confirmed by Google |
| Speedup factors, benchmark scores, success rates | Not confirmed |
| Named researchers or on-record quotes | Not confirmed |
| Exact launch date of the research post | Not confirmed |
| Independent replication of any result | Not confirmed |
None of this makes the announcement empty. A research blog post often precedes detailed papers. Still, readers should treat the reported results as claims from the developer, not as audited findings.
Market Impact: Agentic Coding Gets a Team Layer #
The commercial signal is the more durable part of this release. Google is moving from selling model access to selling coordinated work. A model priced per million tokens competes on price and quality. A team product competes on outcomes, and outcomes are harder to copy.
Pricing hints at the design. Hydrogen’s low listed rate and Argon’s introductory rate both encourage developers to run many agents at once. A team architecture multiplies token use, so cheaper workers are a requirement, not a perk. Whoever controls the cheap worker tier controls the cost of a full /teamwork run.
Enterprise buyers care about a second effect. Repository-scale migrations and multi-milestone projects are exactly the jobs that consume engineering quarters. If a coordinated agent team handles even part of that work under human review, the budget conversation shifts from seats to projects. Google has not published customer numbers, so this remains a thesis, not a measured trend.
The science angle adds prestige and a different buyer. Universities, national labs, and pharma groups buy compute and tools, and a reported track record in chemistry and mathematics gives Google a story that pure coding vendors cannot tell.
Competitive Comparison: How Google’s Stack Stacks Up #
Google is not alone in packaging agents. Every major lab now sells some form of long-running agent, and the fight has moved from single answers to supervised workflows. Our model-by-model comparison of Gemini and Claude pricing across Hydrogen, Fable 5.2, and GPT-6.1 Terra shows how tight the price race has become.
Developer tooling is the other front. GitHub’s agentic coding push, which we covered when Copilot added Carbon and Fable 5.2 on Pro+ and Max plans, shows that Google’s own models now reach developers through rival products as well. Google wants the team layer to live inside Antigravity, but a model sold through a competitor’s editor still earns revenue.
Where does /teamwork differ? Based on the announcement, three things stand out. It leans on formal verification instead of only human review. It names scientific domains, not only software. And it ships as a built-in command inside an existing agent product, which lowers the barrier to a first trial. No competitor comparison on speed or accuracy is possible yet, because Google has published no head-to-head numbers.
By contrast, rivals have tended to market agent products on benchmark scores. Google’s research-first framing is a different bet. It asks the community to judge the system by problems solved with experts, not by a leaderboard rank.
Historical Context: A Month of Gemini 4 Moves #
Read the announcement alongside the rest of Google’s 2026 science and agent work, and a pattern appears. DeepMind’s biology effort, AlphaGenome 2, is the closest precedent, and our coverage tied it to 100M+ cell profiles across 68 tissues. That project aimed a model at a single scientific domain.
The Gemini 4 family widens the aim. Carbon, Argon, and Hydrogen share a design philosophy of deep reasoning paired with cheap execution, and /teamwork is the first public showing of that philosophy applied across several disciplines at once. The step from one specialist model to a coordinated team of general models marks a real change in strategy.
The release schedule supports that reading. Google’s own pages date the Carbon launch to early November and the Hydrogen launch to days later, which means the team product arrived almost immediately after the models it depends on. Shipping the models and the orchestration layer within one week suggests the two were planned together.
Who Benefits First: Researchers, Platform Teams, Kernel Engineers #
Three groups stand to gain soonest. Researchers with formalizable problems come first. A mathematician who can express a conjecture in Lean 4 gets a tireless collaborator whose output a machine can check. That workflow rewards precision and punishes vague goals.
Platform and infrastructure teams come second. Repository-scale migrations, the kind that stall for quarters, suit a scoped, milestone-driven team agent. Those teams already own large test suites, and an unmodified suite is exactly the safety net the announcement highlights.
Performance engineers come third. Kernel optimization has a built-in judge in the form of hardware counters. An agent team can try variants, measure each one, and keep the winners. The human still sets the target and reviews the result, but the search itself gets cheaper.
Teams without clear acceptance tests will benefit least. A task with no automated check leaves the agents guessing, and the human reviewer carries the whole verification burden.
Risks: Verification Gaps, Cost Creep, and Over-Trust #
Multiagent systems create new failure modes. Agents can reinforce each other’s mistakes, a problem that review gates and external checkers try to catch. The announcement’s emphasis on independent tools reads as an admission that self-review is not enough.
Security belongs on the list too. Our report on Microsoft’s finding that fewer than 10 neurons can backdoor AI agents is a reminder that agent pipelines have their own attack surface. A team that reads the web, runs code, and edits repositories deserves the same scrutiny as any privileged automation.
Cost is the quieter risk. Argon’s listed rates double after November 30, and a long team run multiplies every per-token charge by the number of agents involved. Budget caps and checkpoints belong in the plan from day one.
Finally, there is over-trust. A formally verified proof can still answer the wrong question, and a passing test suite can still miss a bug. Verification narrows risk without erasing it, which is why human review stays central.
How to Evaluate /teamwork Before Committing a Project #
Teams interested in a trial can follow a simple routine. Start with a bounded task that has an automated pass or fail signal, then compare the agent output with what your own team would have produced. Per Antigravity’s documentation, the session opens with a scoping interview, so expect to answer questions before any code changes.
A cautious first run
The sketch below is an illustrative prompt we wrote, not a Google example. It shows the shape of a task that suits the tool: a clear goal, a fixed test suite, and a stop condition.
/teamwork migrate the payments module from the old HTTP client
to the new one. Keep the existing test suite unmodified and green.
Stop at the first milestone and report what changed.
What to measure
Track four things: the share of changes that pass your unmodified tests on the first try, the human review time per change, the token spend for the run, and the number of defects found after merge. Without those four numbers, any claim of productivity is a guess. Teams that already log this data for human work can compare directly.
Five Predictions for Multiagent Science Tools #
These are our forecasts, not reported facts. Each rests on the pattern visible in Google’s announcement and the surrounding model releases.
- Verification will become the selling point. Vendors will compete on how rigorously agents are checked, not only on how smart the models are.
- Rivals will ship team commands. Expect competing labs to package orchestration features within months, since the idea is easy to copy even if the results are not.
- Cheap worker models will set the price floor. Hydrogen-class models will define what a full team run costs.
- Independent replication will decide credibility. The first outside groups to publish results with
/teamwork, good or bad, will matter more than Google’s own write-up. - Science buyers will test before they buy. Labs will pilot on formalizable problems first and expand only after seeing audited output.
For broader coverage of Google’s model line and its rivals, see our AI and machine learning section. We will update this analysis if Google publishes figures, papers, or named collaborators.
FAQ: Antigravity’s /teamwork #
What is Antigravity’s /teamwork?
It is a multiagent command inside Google’s Antigravity product. Per Google Research, it supports human-AI collaboration on open problems in mathematics, chemistry, systems software, and ML engineering. Per Antigravity’s documentation, it targets large software projects, simulations, and deep research.
Which models does it use?
Google names Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen. Carbon is the strongest reasoner, Argon is the efficient frontier model, and Hydrogen handles fast execution as a worker or verifier.
How is /teamwork different from /boost?
The /boost command runs an on-demand multi-agent reasoning pipeline for hard engineering tasks like bugs and refactors. The /teamwork command coordinates a longer project with planning, roles, and review gates.
How does Google verify the agents’ work?
The announcement names the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, and hardware counters. It also says the work was done with human experts.
Are there benchmark numbers or speedups?
Not in the source material we reviewed. Google Research reports application areas such as kernel optimization and molecular-structure discovery without attaching figures in that material.
What do the Gemini 4 models cost?
Google lists Hydrogen at $1.50 per 1 million input tokens and $7.50 per 1 million output tokens. Argon is listed at $2 and $10 through November 30, rising to $4 and $20 afterward. Google has not published a price for /teamwork itself in the material reviewed.
Can outside developers use it today?
Google says the tools are being made available externally through /teamwork, alongside /boost and other agents. Availability can vary by product and account tier, so check your Antigravity account for access.