The development

What happened

Google Research has detailed how Antigravity’s /teamwork combines AI agents with external checks for research and engineering. Its 13 November update reports new applications, while distinguishing tools available now from specialized workflows still to come.

Key facts

  • Humans frame problems and constraints; agents explore candidates while checks such as Lean, unmodified test suites and hardware evaluations assess results. Separate contexts support specialist workers and fresh critics.

  • In its DuoDiscover example, Google reports 4.09-times geometric-mean speedup across 17 JAXBench kernels against reference implementations. Fast proxy evaluations are periodically checked against full hardware runs.

  • Google says /teamwork, /boost and Deep Review are available externally. The formal-proof pattern and additional specialized workflows are forthcoming; applications for six months of high-tier access target 10,000 academic researchers, not every user.

The context

Why it matters

TideReach assessment: an external evaluator can make an agent’s candidate easier to test, but the evaluator’s scope still matters. These are Google-reported results, not TideReach reproductions or guarantees for other tasks. The update does not mean every described workflow has shipped.