Agents Enter Workflows. Infrastructure Hardens. Physical AI Scales.

What happened
Notion expanded beyond note-taking and launched a new developer platform, adding cloud-based “Workers” for custom code, database sync for any API-backed data source, and support for external agents such as Claude Code, Cursor, Codex, and Decagon inside the Notion workspace. Notion says customers have already built more than one million custom agents since the February launch.

Why it matters
This is a shift from “AI features inside productivity software” to a bid for orchestration control. Notion is trying to become the place where company data, internal workflows, and third-party agents meet — a much more valuable position than just being another app with a chatbot.

What’s next
The big question is whether teams really consolidate agent workflows inside Notion instead of keeping them scattered across scripts, integration tools, and separate agent products. Notion making Workers free through August suggests it is prioritizing developer adoption and ecosystem lock-in first.

What happened
OpenAI published a detailed engineering post on a new Windows sandbox for Codex, explaining that it moved from an earlier “unelevated” prototype to an “elevated sandbox” design using restricted tokens, dedicated local users, and firewall rules to better constrain file writes and outbound network access. The company said the change was needed to make Codex “just as safe and delightful to use on Windows” as on other operating systems.

Why it matters
This is the unglamorous work that makes agentic software believable in real organizations. If coding agents are going to run on employee laptops and touch files, branches, package managers, and local tools, trust will depend as much on sandboxing and permissions as on model quality.

What’s next
Expect more competition around runtime governance, approval flows, and machine-level controls as agentic coding tools spread. The likely next battleground is not just who has the smartest agent, but who can let enterprises deploy one safely on the systems people actually use.

What happened
Ramp’s latest AI Index shows Anthropic overtook OpenAI for the first time among participating businesses, with 34.4% paying for Anthropic services versus 32.3% for OpenAI; Ramp says its sample covers more than 50,000 American businesses using corporate spend data. On the same day, Axios reported that Anthropic also launched “Claude for Small Business,” connecting Claude to tools including QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365.

Why it matters
This is a strong signal that the competitive fight is no longer just about benchmarks or consumer mindshare. Anthropic appears to be winning with workflow fit and technical buyer appeal, then using that momentum to expand into small businesses — a market bigger in count, messier in execution, and harder to serve well.

What’s next
If Anthropic can turn early SMB interest into durable usage while holding its lead with technical teams, the OpenAI versus Anthropic contest will look increasingly like a distribution and packaging war, not just a model war. OpenAI still has broad reach, but the enterprise leaderboard is no longer static.

What happened
NVIDIA announced an engineering collaboration with David Silver’s Ineffable Intelligence to build reinforcement-learning infrastructure at scale, starting on Grace Blackwell systems and exploring the upcoming Vera Rubin platform. NVIDIA framed the effort around “superlearners” — systems that learn continuously from experience instead of relying only on fixed human-created datasets.

Why it matters
That framing matters. It suggests a growing belief that the next big capability gains may come less from feeding models more static internet-scale data and more from online learning loops, simulation, and experience-driven training. In plain English: the frontier may be shifting from bigger models to better learning systems.

What’s next
Watch for more labs and infrastructure vendors to talk about serving loops, interconnect, memory bandwidth, and simulation pipelines instead of only pretrained parameter counts. If reinforcement learning becomes the next serious scaling vector, the infrastructure stack around it becomes strategic again.

What happened
TechCrunch reported that Rivian spinoff Mind Robotics raised another $400 million on May 13, just two months after a prior $500 million round. The company’s new funding brings total capital to more than $1B and values it at more than $3B, while its own announcement says it is building an industrial robotics platform that combines foundation models, hardware, and deployment infrastructure for live manufacturing environments, with Rivian acting as a key partner. The Wall Street Journal separately reported a $3.4B valuation and plans to train and deploy hundreds of robots at Rivian’s factory in Normal, Illinois.

Why it matters
This is one of the cleaner signals that physical AI investment is moving toward deployment narratives, not just humanoid spectacle. Investors are backing a company that pairs model development with a real factory loop — the kind of setting where robots can generate the data, feedback, and iteration cycles that physical AI needs to improve.

What’s next
The crucial test is whether Mind can turn Rivian access into repeatable deployments outside its home turf. If it can, manufacturing could become one of the first sectors where physical AI develops a true adoption flywheel rather than living as a perpetual pilot.

What happened
Humanoid announced a major deployment partnership with Schaeffler that will put thousands of humanoid robots into real manufacturing environments over time, marking one of the largest industrial humanoid rollouts announced so far. The deal signals a shift from pilot programs toward production-scale deployment in automotive manufacturing.

Why it matters
The humanoid robotics market has been full of flashy demos, but very few companies have secured enterprise-scale manufacturing commitments. Schaeffler isn’t experimenting with a showroom robot — it’s betting that humanoids are becoming economically viable for repetitive industrial labor.

What’s next
The real test now is operational reliability: uptime, safety, maintenance, and labor economics. If humanoids can consistently handle factory workflows at scale, manufacturing could become the first true breakout market for embodied AI — and trigger a broader deployment race across logistics, warehousing, and industrial automation.

AI’s next phase is shifting from chat interfaces to operational infrastructure. The winners may not be the companies with the smartest models alone, but the ones that control where agents run, how safely they operate, how continuously they learn, and how deeply they integrate into real workflows, factories, and enterprise systems.

Build your own lightweight “agent workspace” stack. Use Notion with Cursor or Claude to create an internal research or operations agent that pulls live API data into a shared workspace. Then experiment with approval flows, sandboxing, and permissions — the same governance problems enterprises are now racing to solve for production agents.

For a second layer, explore how reinforcement learning and simulation may shape the next generation of AI systems. Study how NVIDIA is positioning infrastructure for continuous-learning “superlearners,” then compare that to how physical AI companies like Humanoid and Mind Robotics are using real-world factory environments as training loops for embodied agents.