Field noteNov 16, 20265 min read1 source

Google Research published its own version of this Friday.

In one minute

Google Research says Antigravity’s /teamwork multi-agent workflows are now available to external researchers and developers, alongside Gemini 4 and /boost. The announcement combines a newly described research workflow with claims across mathematics, chemistry, and software engineering; its broader significance is a practical architecture built around specialized agents and independent verification.

What changed: Google Research describes Antigravity’s teamwork feature applied to proving the Courtade-Kumar conjecture, adding a mathematical-science use case for the agent platform.

What happened

In a November 13, 2026 announcement, Google Fellow and Google Research VP Vahab Mirrokni described Antigravity’s /teamwork as a framework for long-running collaborative work. Google says researchers frame questions and constraints, while agents write solvers, explore candidate designs, and check reasoning. Its design addresses two limitations: unpredictable model generations and loss of detail over multi-day jobs. Parallel exploration is paired with dedicated falsifiers such as the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, and hardware counters. Context is partitioned so each specialized agent receives only the files and tools needed for its step; coordinators pass work to fresh successors, while critics and auditors review outputs from clean contexts.

Google says users can express workflows as declarative natural-language patterns, which route prompts to a suitable harness. These include distributed coding and iterative problem solving, packaged as /boost, as well as research workflows for long proofs and document review. The company says Gemini 4 models with /teamwork and /boost are usable in Antigravity today. It also opened applications for 10,000 academic researchers to receive six months of free access to its highest usage tier. Google says a formal-proof pattern and additional specialized workflows are planned for later updates, with tutorials to follow.

The announcement’s mathematical example concerns the Courtade–Kumar conjecture, an information-theory problem posed in 2013. Google says researchers from Google, CUHK, USC, and CMU used the Long Proof pattern to prove it in full generality. Gemini first found counterexamples to a generalized conjecture from earlier CUHK work. The CUHK researchers proposed a candidate function that reduced the problem to a 14-variable optimization problem and supplied an analytical lower bound. Gemini 4 Argon and earlier Gemini models then partitioned the parameter space, developed most of the new analytical bounds and interval-arithmetic certificates, and extended the proof from balanced to unbalanced functions without human input. The team formalized the proof in Lean 4; Google reports a 50-million-line formalization.

Other examples illustrate the same agent-and-verifier approach. For determining molecular structures from nuclear magnetic resonance data, Google says 30 specialized agents and a physics-informed simulator achieved 95% top-1 and 100% top-3 accuracy on complex benchmark molecules, nearly doubling single-agent performance. For a Rust rewrite of Google’s libgav1 decoder, the company reports bit-exact parity across a broad conformance suite and, after profiling-based optimization, performance 20–30% faster than the original C++ and hand-optimized assembly on x86_64 and ARM. Google also describes production infrastructure work, including a Borg scheduling change associated with 24% fewer TPU workload preemptions on some hardware types while preserving service-level objectives and cluster utilization.

What the archive adds

New Runtime’s earlier coverage includes Antigravity’s antimicrobial-peptide search workflow on October 21, 2026, and an Antigravity cheminformatics workflow on November 3, 2026. This is a new event: Google is describing /teamwork’s broader research and engineering workflows and opening access, not merely repeating those scientific applications.

Trend signal

This supports harness architecture outlives model choice. Google’s account puts context partitioning, task routing, specialized roles, and external checks at the center of the workflow, even as multiple Gemini versions supply the models. The announcement presents the harness as the mechanism that turns model output into work that can be checked, rather than treating model identity alone as the system.

Why now

Google says external scientists and developers can use the tools today, while applications are open for a limited cohort of 10,000 academic researchers to receive six months of its highest usage tier. That makes the announcement an access change as well as a collection of internal and collaborative demonstrations.

Why it matters

The most reusable idea is not that agents can produce an unusually long proof or port a large codebase. It is the division of labor around verifiable outputs: specialists explore, separate reviewers challenge their work, and objective tools decide whether a candidate passes. That architecture gives teams a way to spend human attention on framing constraints and checking whether formalized claims match the original problem, rather than trusting a fluent answer.

The concrete examples also show where this approach is strongest: problems with an evaluator, such as a proof checker, test suite, simulator, or hardware measurement. But Google’s impressive results are demonstrations, not a general guarantee that multi-agent systems will work reliably on less measurable tasks. Builders should treat the published workflow as a design pattern to test against their own tasks and verification budget, not as evidence that adding more agents automatically improves output.

Limits and uncertainty

The announcement is Google’s account of its own research and systems work. The results described are not independently assessed here, and the excerpted claims do not establish reproducibility across teams, prompts, or different hardware. In particular, a 50-million-line Lean formalization does not by itself establish that the theorem statement matches the intended informal conjecture; Google notes that human mathematicians confirm that correspondence. It is also unclear how access limits, usage quotas, and the application process work beyond the stated cohort and duration. The company says more workflows are forthcoming, but does not give a schedule in this announcement.

Skill to build now

For a task with an objective pass/fail check, write down the evaluator before assigning work to agents. If it is a code change, preserve the original test suite and require every module to pass it; if it is a formal proof, define the theorem statement for human review and require Lean to check the result. Then separate generation from criticism: give reviewers the artifact and criteria, not the producer’s full reasoning, so they can look for failures independently.

Reader takeaway

Antigravity’s /teamwork is a newly accessible example of agent orchestration organized around independent verification. The practical lesson is to start with a task whose result can be checked by an external tool, then design agent roles and handoffs around that check—not to assume that parallel agents alone make difficult work trustworthy.

Evidence / sources

  1. 1research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicGoogleExplore the google topic hub.
  2. 02topicGeminiExplore the gemini topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract