Agents are doing the work now: a proof, a campus and a hospital ward

A proof, a kernel, and a review queue

Google Research released Antigravity's /teamwork, a multi-agent system for human-AI collaboration and designed for long-horizon problems in mathematics, chemistry, and systems engineering. The system uses parallel exploration paired with dedicated falsifiers to ensure reliability, and it partitions context by construction to prevent attention dispersion. Researchers used /teamwork to prove the 2013 Courtade-Kumar conjecture in information theory, and the /teamwork-LEAP framework produced a machine-checked formalization of the Weil conjectures. A machine-checked formalization is the key phrase: the result is not a model claiming a proof, but a proof a machine has verified. m:…9331

The engineering result is just as striking. A two-week Argon campaign rewrote the Zircon microkernel in Rust, passing all unit tests. Deep Review provided feedback on 578 submissions for STOC 2027, with 90 percent of surveyed authors rating it as factual. DuoDiscover achieved a 4.09 times geometric-mean speedup over reference implementations across 17 JAXBench kernels, and Borg's agentic scheduling optimization resulted in a 24 percent reduction in TPU workload preemption. Google is offering 10,000 academic researchers six months of free access to the highest usage tier of Antigravity tools, which suggests the company wants the research community to test the system on problems it has not chosen itself. m:…9331

Career services without new headcount

The University of Colorado Colorado Springs faced the most ordinary problem in this set: growing demand for career support without the capacity to add career center staff. Career readiness is a core goal in the university's strategic enrollment plan, and its answer was an AI career services agent named Journey. Journey handles help desk questions, resume reviews, and mock interviews, and is integrated into the Gateway Program Seminar at UCCS, so students meet it in their first year rather than discovering it in their last. m:…8655

The conversation mix shows what students actually ask. Job and internship questions account for 25 percent of Journey's conversations, career center logistics account for 25 percent, major and career exploration account for 20 percent, resume and cover letter review account for 10 percent, and Clyde's Closet accounts for 10 percent. Journey provides 24/7 mock interview availability and coaching summaries via text, and directs students to human career coaches. The design is the point: the agent handles the routine volume so the humans can do the high-touch counselling, and it hands off rather than pretending to replace them. m:…8655

Chart review inside the firewall

Health systems and startups are using AI to support clinicians and streamline workflows, and Dyania Health built Synapsis AI to analyze decades of patient records. Synapsis AI works across notes, reports, and lab results to connect findings, runs on NVIDIA-accelerated computing inside a health system's firewall, and patient data processed by Synapsis AI does not leave the hospital. Dyania Health uses NVIDIA NeMo open libraries to build and train clinical models, and NVIDIA TensorRT and TensorRT-LLM to optimize models for GPU servers. m:…3674

The Cleveland Clinic study is the evidence. Cleveland Clinic used Synapsis AI to support a study of over 7,800 aortic valve replacement patients, covering patients between 2005 and 2021, and the AI platform completed chart review for the cohort faster than manual methods. Synapsis AI built a predictive model of valve deterioration and reintervention risk, covering the three-, five-, and ten-year marks post-procedure. A predictive model over that cohort, built from records that never left the hospital, is the kind of deployment that survives scrutiny precisely because it was built inside the constraints rather than around them. m:…3674

The through-line across all three is verification. The research system pairs exploration with dedicated falsifiers. The campus agent directs students to human career coaches. The hospital AI works inside the firewall on data that never leaves. None of these deployments asks anyone to trust the agent. Each one builds the check into the workflow, and that is why they are working while so many agent pilots are still demos. m:…9331 m:…8655 m:…3674