antigravity teamwork: Google's Multi-Agent System Opens to Everyone

antigravity teamwork, Google's multi-agent orchestration feature, crossed an important line on November 13: Google Research published a results post showing what the system has done on real open problems, and said external scientists and developers can now use /teamwork and /boost in Antigravity with Gemini 4 models, with applications open for 10,000 academic researchers to get six months of free access to the highest usage tier. Source

What changed on November 13

Teamwork itself is not new: Google first announced the framework at Google I/O and shipped a major update in August, when agents learned to propose, critique, and refine each other's work autonomously over hours or days. The November 13 post changes two things. First, Google published a stack of attributed results across mathematics, chemistry, and systems engineering, all produced by human researchers working with /teamwork harnesses on Gemini 4 Argon, Carbon, and Hydrogen. Second, the tools move from internal and partner use to general external availability, alongside Deep Review, with more specialized patterns promised in the coming weeks. Source Source

The results Google is claiming

The headline math result: researchers from Google, CUHK, USC, and CMU used /teamwork's Long Proof pattern to prove the Courtade-Kumar conjecture, open since 2013, in full generality. Google says Gemini first found counterexamples to a generalized conjecture from the CUHK team's earlier paper, and the collaboration converged on a full proof from there. The same post reports work in optimization, statistics, quantum computing, algorithms, molecular structure discovery, and a generated causal inference benchmark. Source

The systems numbers are the ones engineers will quote. Google reports FarArray, a kernel-bypass partitioned global address space built with /teamwork and Argon, whose lock-free Connected Components implementation scales to 1-trillion-edge graphs with 60x to 90x throughput gains at 6x lower RAM over TCP key-value stores; an InstantEmbedding workload that cut peak DRAM by up to 14x; and custom TPU kernels for graph neural networks delivering up to a 185x high-bandwidth-memory reduction. A two-week campaign rewrote the remaining C++ of the Zircon microkernel into about 800,000 lines of safe Rust that pass all kernel unit tests and boot on arm64 hardware, and a separate effort ported libgav1, Google's AV1 decoder, to roughly 170,000 lines of Rust with 100% bit-exact parity across a conformance test suite. In a pure long-horizon test, a hill-climbing team trained a 9x9 Go engine from a basic AlphaZero implementation to 1,885 Elo in under eight days. All of these are Google's own reported figures. Source

How /teamwork actually works

The design addresses two known failure modes of long agent runs. First, single-shot generation is unreliable, so the system pairs parallel exploration with dedicated falsifiers: the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, or hardware counters automatically verify which candidates work. Second, models lose track of details over multi-day tasks, so /teamwork partitions context by construction: each specialized agent receives only the exact files and tools it needs, coordinators hand off to fresh successors between milestones, and critics and auditors verify results from clean contexts without the original producer's reasoning. Workflows are expressed as declarative natural-language patterns, and the system routes each prompt to the pattern that fits. Source

The docs spell out the org chart. Sentinel coordinates once you approve the brief, spawns a Project Orchestrator that decomposes the work into milestones and hands off to a fresh successor between them, Explorers do read-only research, and Workers implement in non-overlapping tracks. Before a milestone is approved, candidate changes pass three adversarial gates: a Critic doing independent code review, a Challenger building adversarial test suites and worst-case inputs, and an Auditor that checks test evidence against real command output so tests genuinely passed rather than being mocked or skipped. A Success Auditor runs the final end-to-end pass before anything is presented back. Everything runs inside isolated project directories with artifact handoffs instead of a growing shared context. Source

Execution paths adjust the team to the task: a General path for distributed coding and research, a lighter Iterative coding path for self-contained changes, Document Review for multi-angle critique of papers and design docs, and Math/Proof paths including a large-team tournament mode for hard conjectures. Every project starts with a scoping interview phase before any code is written. The /teamwork command is available on paid plans in both the Antigravity app and the CLI. Source

The two-phase workflow

Every /teamwork project runs in two phases. Phase 1 is a scoping interview: the main Antigravity agent asks what you want built, its purpose, and its audience; drafts requirement blocks limited to what you care about; and, most importantly, agrees with you on an independent verification for each requirement, such as a test suite, a reference benchmark, or an agent judging against an explicit rubric, plus acceptance criteria and a dedicated workspace directory. You review the resulting prompt artifact and approve with a single confirmation. Phase 2 is autonomous execution: Sentinel hands the approved brief to the Project Orchestrator, and the team coordinates through structured artifacts in the workspace, a request artifact, a project plan, and progress records, instead of chat history. The design goal is alignment up front and no constant back-and-forth during the run. Source

What is available now, and what is coming

Per the research post, external users can use Gemini 4 models with /teamwork and /boost in Antigravity and Deep Review today; /boost packages a sequential solve-and-verify loop refined from prior trajectories as a standalone command. The academic offer is 10,000 researchers, six months, highest usage tier, by application. Still to come: the formal proof pattern (the Long Proof pattern behind the Courtade-Kumar result, currently branded Stellar Colosseum internally per the post) and additional specialized workflows, with deep-dive tutorials promised on the Antigravity blog. Independent coverage at Agentic Weekly confirms the November 13 timing and the 10,000-seat offer, while flagging that the headline results are vendor-reported and not independently tested. Source Source

Our read

This section is our assessment, not Google's claims. The results are impressive and unaudited, so treat the numbers as vendor benchmarks until outside teams replicate them. But the architecture is the real story, and it is public in the docs. Three ideas here match what we see working in production agent systems. Clean-context verification, where a Critic and Auditor review work without inheriting the producer's reasoning, is the strongest answer yet to agents grading their own homework, and it mirrors the independent-signal discipline in our AI agent regression testing post. Context partition by construction, giving each agent only the files and tools for its step, is the same least-privilege instinct applied to attention, and it sidesteps much of what goes wrong in long async subagent runs. And the Auditor's insistence on real command output over claimed test results should be a standard feature everywhere; any team asking how many tools an AI agent should have should also ask who checks the receipts.

The practical catch is cost and lock-in. Multi-agent teams with adversarial gates burn tokens by design, and /teamwork sits on paid plans inside Google's IDE and CLI, so the patterns are portable but the implementation is not. If you run agent workflows elsewhere, the transferable lessons are the falsifier pairing, fresh-successor handoffs, and adversarial audit gates, not the slash commands. Watch for the promised tutorials before betting a workflow on the exact role split.