Advancing science and engineering with Antigravity's /teamwork

Google launched external availability for its Antigravity /teamwork multi-agent orchestration system on November 13, 2026, pairing specialized AI harnesses with Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen to tackle longstanding problems in science, mathematics, and software engineering. According to research.google's report by Google Fellow and Vice President of Google Research Vahab Mirrokni, the release follows weeks of joint work with academic institutions to solve open theoretical and computational challenges that had remained unresolved for decades.

This report details the architectural mechanisms behind the /teamwork harnesses, the mathematical conjectures resolved during internal validation, and the concrete algorithmic benchmarks achieved across research domains. The analysis provides systems engineers, theoretical researchers, and machine learning practitioners with technical specifications, verification requirements, and deployment limits for orchestrating long-horizon reasoning agents.

01Partitioned context and automated falsifier loops

The system addresses two structural failure modes in current frontier language models: output unpredictability during single attempts and context degradation across multi-day tasks. To prevent reasoning drift, the platform isolates context windows so specialized agents only ingest files and tools required for their assigned subtasks. Independent coordinator agents then transfer intermediate state to fresh successor instances without carrying forward accumulated reasoning history.

Parallel exploration tracks pair with automated verification tools to filter speculative outputs against concrete computational boundaries. The platform deploys the Lean 4 proof checker, unmodified software test suites, execution simulators, hardware performance counters, and web retrieval as dedicated falsifiers. Candidate solutions that fail environment checks are discarded before downstream agents consume the results.

02Fifty million lines for Courtade-Kumar

Researchers from Google, CUHK, USC, and CMU deployed the Long Proof harness pattern, implemented as Stellar Colosseum within Antigravity, to resolve the 2013 Courtade–Kumar conjecture in full generality. The information theory problem posits that when a random bit string traverses a noisy channel, no single-bit compression scheme retains more mutual information than a dictator function that simply outputs a single raw input bit.

Gemini first discovered counterexamples to an earlier generalized conjecture proposed by the CUHK team, prompting collaborators to reformulate the problem into a 14-variable optimization task with an analytical lower bound. Gemini 4 Argon partitioned the remaining parameter space, derived analytical bounds, generated interval-arithmetic certificates, and extended the proof to unbalanced functions without human intervention. The complete formalization in Lean 4 produced 50 million lines of verified code.

03Quantum queries and matrix multiplication bounds

The multi-agent infrastructure settled a 2014 open problem in adaptive data analysis by proving and Lean-verifying near-optimal sample bounds for statistical queries selected adaptively from previous responses. The derived mathematical bounds produced the first nearly space-optimal streaming algorithm capable of handling arbitrary stream insertions and deletions while maintaining robustness against adaptive adversaries.

In quantum complexity, the system resolved a separate 2014 question by constructing the first n-bit Boolean functions requiring strictly more than n/2 quantum queries at constant error while establishing exact query overhead for symmetric functions. For combinatorial algorithms, a new deterministic construction resolved the central open problem from STOC 2024 by breaking the fifty-year barrier for Boolean matrix multiplication and triangle detection without algebraic cancellations.

For high-dimensional geometry, the agents established a lower bound proving that approximating all pairwise distances across n points within a constant factor can demand polynomially more than n edges, even when utilizing extra Steiner points and unbounded hops. In randomized linear algebra, stacked sparse sketches settled a 2025 question by achieving matrix compression to linear dimensions using only the square root of log r nonzeros per row.

04Integer arithmetic and the Deque conjecture

For dynamic weighted graphs, the Stellar Colosseum harness resolved an open question from FOCS 2022 by introducing the first single-pass cut sparsification streaming algorithm that simultaneously minimizes memory, update latency, and reconstruction time. In optimization theory, the system resolved an open problem in online inverse linear optimization by generating a deterministic algorithm that delivers optimal regret while reducing runtime from exponential to polynomial complexity.

The platform also produced a new O(n log n) integer multiplication algorithm operating on a single two-dimensional grid whose C implementation matches the GNU Multiple Precision Arithmetic Library bit-for-bit up to one billion bits. A separate run achieved progress on Tarjan's 1985 Deque Conjecture for self-adjusting splay trees, establishing an upper bound that surpasses every finitely iterated inverse-Ackermann function, verified directly inside Lean 4.

In theoretical statistics, the multi-agent system established that approximate leave-one-out cross-validation maintains mathematical accuracy when applied to non-smooth sparse penalty functions like elastic net. This validation confirms that researchers can bypass repeated model refitting without sacrificing error bounds during high-dimensional data evaluations. Across integer multiplication, multi-agent runs also streamlined classical theoretical bounds on Turing machines below n log n.

05Declarative harness patterns and researcher access

Workflows in the system run as declarative, self-improving natural-language templates that dynamically route incoming prompts to specialized harnesses based on task characteristics. These operational topologies range from distributed code generation and document analysis to iterative problem-solving loops packaged as the standalone /boost command. Teams can also build custom specifications by extracting coordination logic and audit rules from prior execution transcripts.

Google is distributing these agentic capabilities through the Antigravity platform alongside a dedicated researcher access program for external institutions. The workflow includes critic and auditor agents that examine generated artifacts from clean contexts without access to prior scratchpads. While core reasoning features and /boost are active today, Google plans to release additional harness configurations in upcoming weeks.