
Agentic AI
🧞 Databricks Makes Genie an Always-On Coworker
What happened
Databricks introduced a second generation of Genie designed to run continuously in the background, manage long-running work, delegate to subagents, follow up on missing information, and surface decisions through a priority inbox and email. Genie Ontology is generally available today, while the new always-on Genie, Genie Search, and Ontology Manager are coming to preview and beta programs.
Why it matters
The enterprise agent is starting to look less like a chatbot and more like a persistent worker with goals, business context, permissions, and its own compute. Databricks is betting that autonomy becomes more useful when the agent understands how the company defines metrics, people, processes, and access.
What’s next
Customers will need to test whether long-running agents remain reliable as tasks stretch across multiple systems and decision points. Unity Gateway and Unity Catalog provide controls around tool access, usage, cost, and permissions, but the always-on version itself has not yet reached general availability.
👥 Anthropic Finds More Agents Can Mean Worse Outcomes
What happened
Anthropic researchers tested frontier models from Anthropic, OpenAI, and Alibaba in four simulated environments where several users delegated to separate agents that had to share resources. In every environment, group outcomes were worse when each user had their own agent than when one agent coordinated for everyone.
Why it matters
Multi-agent systems create a coordination problem that better individual reasoning does not automatically solve. Researchers observed agents stalling as teams grew, overriding other users’ requests, and making unsupported claims about other people.
What’s next
Anthropic released MAMUBench so developers can test these failure modes themselves. The research also found that explicit leadership, clearer priorities, and rules requiring agents to read messages before acting improved outcomes in some environments.
📱 BotPsy Shrinks Tool-Using Agents to 0.8B Parameters
What happened
Tether AI Research released BotPsy, an open-weight 0.8-billion-parameter model designed specifically to call tools and execute multi-step actions on phones and laptops. Tether says BotPsy outperformed every model of 1.2 billion parameters or fewer that it tested across six agent benchmarks and exceeded larger models on several tool-use evaluations.
Why it matters
Personal agents need access to highly sensitive context such as calendars, messages, files, and contacts. Running the decision model locally could reduce cloud dependence, latency, and exposure of personal data while making agentic capabilities practical on smaller devices.
What’s next
The weights are available under Apache 2.0 for developers to download and integrate, but BotPsy is a model rather than a finished consumer app. Tether’s benchmark results remain vendor-reported and will need broader independent testing across real multi-step workflows.
Generative & Enterprise AI
🗄️ Google Brings AI Closer to Enterprise Data
What happened
Google Cloud added new AI capabilities to AlloyDB that can automatically classify text and pull structured information from documents and records. Instead of sending data to a separate AI application, companies can run those tasks closer to where the information already lives.
Why it matters
Enterprise AI often gets complicated because data has to move between databases, applications, and models. Google is trying to remove some of that friction by bringing AI directly into the systems companies already use to store operational data.
What’s next
The new capabilities are still in preview. Google plans to add more cost controls, better tracking of where AI-generated information came from, and additional ways to run AI directly against enterprise data.
⚡ Torus Brings AI Compute to Where Power Already Exists
What happened
Torus launched Foundry, a modular AI infrastructure system combining power, storage, cooling, networking, and compute at existing commercial or industrial sites. The company says deployments can range from 1 to 20 megawatts and move from contract to compute in 12 to 16 weeks, and it announced $500 million in new growth capital alongside the launch.
Why it matters
AI infrastructure is increasingly constrained by power and grid access, not only by GPU supply. Torus is trying to move inference capacity toward existing electrical headroom instead of waiting for entirely new hyperscale campuses to be built.
What’s next
Torus says thousands of GPUs are already operating in Foundry units and targets more than 47,000 by the end of 2027. Those deployment figures and timelines are company claims that will become more meaningful as additional customer sites come online.
Physical AI
🚚 Serve Opens Its Robot Fleet to Developers
What happened
Serve Robotics launched the AV Delivery Developer Portal, exposing a public API and sandbox that let developers, businesses, and AI assistants build workflows around its fleet of more than 2,000 sidewalk delivery robots across over 150 U.S. neighborhoods. Applications can request a pickup and dropoff, confirm details, and track a robot while Serve handles dispatch, navigation, and supervision.
Why it matters
Physical AI is becoming programmable infrastructure. Developers no longer need to build a robot or autonomous-navigation stack to create an application that can move something through the real world.
What’s next
The simulated sandbox is available to builders now, while access to live robots requires Serve approval and remains limited to active service areas and operating hours. The bigger question is what new businesses emerge once physical delivery becomes callable like a software API.
🔮 NTT Lets Robots Preview Failure Before Moving
What happened
NTT developed a physics-based world model that evaluates several candidate robot actions before execution and predicts outcomes such as collisions, dropped objects, or collapsing loads. In a small-robot shelf-loading experiment, NTT reports the approach increased task success from 73.1% to 92.3%.
Why it matters
Most robots learn whether an action worked after they take it. A system that can simulate several plausible futures first could make physical AI safer and more adaptable in environments where mistakes damage goods, equipment, or people.
What’s next
NTT plans field trials with Mujin focused initially on loading and unloading varied airport baggage, followed by potential expansion into logistics, manufacturing, construction, and retail. The current results remain research-stage performance rather than proof of production-scale reliability.
💡 Bottom Line
AI is moving from answering requests to owning work for longer periods and across more environments. That makes context, coordination, local execution, infrastructure, and simulation increasingly important parts of the autonomy stack. The next leap will come from agents that can act continuously without losing sight of boundaries, other agents, or the physical consequences of their decisions.
⚙️ Try It Yourself
Dispatch a Robot Without Owning One
Serve Robotics launched a developer sandbox today that simulates the same pickup-to-delivery workflow used by its physical robot fleet. Live robot access requires approval, but Serve says anyone can build and test in the sandbox first.
Try this: Open the Serve AV Delivery Developer Portal and use the sandbox to create a simulated pickup and dropoff. Follow the workflow through confirmation and robot tracking, then change one detail and run it again.
Which parts of the task are exposed to your application, and which responsibilities remain inside Serve’s autonomy stack?
Insight: Physical AI becomes much easier to adopt when businesses can call a robot like an API instead of building robotics infrastructure themselves.
