MongoDB’s Atlas Agent Engine Makes Context a Shared Asset for Agents
Last week I was at MongoDB.local NYC, where the company kept its launch schedule despite CEO CJ Desai leaving for Meta just days earlier. Of the three launches from the event, Atlas Agent Engine matters most. It’s a runtime and governance stack, now in public preview, that turns agent memory and context into a shared asset. That memory lives in the customer’s own MongoDB instance and can run on any major cloud, on-premises, or locally. Decoupling memory from any single model or agent makes it easier to swap models and chain smaller agents together. MongoDB’s own proof point is Holly, a multi-agent sales assistant that is now live for all of its sellers. In my new Field Note, I explain why this is one of the more promising agent infrastructure previews I’ve seen this year. I also cover what MongoDB still needs to prove on partnering, platform scope, and integration as it moves toward general availability.

IBM Bob Goes Local
When I sat down with Kaya Gangesan, a software program director at IBM, at TechXchange last fall for a Six Five demo of what was then Project Bob, getting started was as simple as downloading the IDE, entering an API key, and going. That simplicity is great for developers, but it’s also a reminder that the intelligence lived in IBM’s cloud. That’s what makes this week’s news significant. IBM has made Bob available as a self-hosted offering, and the option supports on-premises, private-cloud, sovereign-cloud, and air-gapped environments, with customers running supported models they’ve licensed locally or connecting to external model services through hybrid configurations.

IBM is extending the same option to the Bob Premium Package for Z, and that’s where the real story is. The code that most needs modernizing in banks, governments, and insurers is the code least likely to be allowed near a cloud-hosted AI service. It’s worth being precise about what “on-premises” means here, because the industry tends to blur two very different things. Nearly every AI coding tool is “locally installed” in the sense that the IDE or CLI runs on the developer’s machine, but the agent’s reasoning, the model inference, and often the repository context still travel to the vendor’s cloud. Running on-premises means the entire stack, including model serving, the agent runtime, and the governance and policy controls, sits on infrastructure the customer owns and operates. In the air-gapped case, it never touches the internet at all. That’s the difference between a tool a regulated enterprise can pilot and one it can actually put into production.

Oracle Enables Autonomous Agents with Its New Fusion Claw Offering
Oracle’s new Fusion Claw is another sign that the agent conversation is moving from assistants to harnesses built for long-running, autonomous work. Oracle describes Fusion Claw as a governed agentic execution runtime that pairs AI reasoning with deterministic enterprise computation, and it is launching with 25 Claw-powered agentic applications. The design choice I find most interesting is the division of labor. A frontier model handles reasoning, planning, and adaptation, then hands off to deterministic computation for high-volume execution, so expensive model calls are used only where intelligence is actually needed.

Governance is equally central. An “Outcome Trust Harness” enforces each organization’s policies, permissions, and approval boundaries on every run, and an “Outcome Receipt” leaves an auditable record of the decisions made and transactions executed. Customers can set each process anywhere from light assistance to fully automated execution within delegated authority. This lines up with what I wrote in last week’s newsletter about Microsoft’s new Copilot client, which includes Autopilot, an always-on agent that runs long tasks under its own identity.

Oracle’s product name may invite comparisons to OpenClaw, but the more important parallel is architectural. Vendors increasingly agree that the model is only part of the product. The harness around it, covering identity, policy, replanning, cost control, and audit, is what makes multi-hour autonomous work something an enterprise will actually allow. The difference is in the starting point. Microsoft is building a horizontal harness for knowledge workers. Oracle is building one tied to the system of record, which narrows the scope but gives its agents the business context and transactional authority that a general-purpose agent has to borrow.

MongoDB’s Atlas Agent Engine Makes Context a Shared Asset for Agents
Last week I was at MongoDB.local NYC, where the company kept its launch schedule despite CEO CJ Desai leaving for Meta just days earlier. Of the three launches from the event, Atlas Agent Engine matters most. It’s a runtime and governance stack, now in public preview, that turns agent memory and context into a shared asset. That memory lives in the customer’s own MongoDB instance and can run on any major cloud, on-premises, or locally. Decoupling memory from any single model or agent makes it easier to swap models and chain smaller agents together. MongoDB’s own proof point is Holly, a multi-agent sales assistant that is now live for all of its sellers. In my new Field Note, I explain why this is one of the more promising agent infrastructure previews I’ve seen this year. I also cover what MongoDB still needs to prove on partnering, platform scope, and integration as it moves toward general availability.

IBM Bob Goes Local
When I sat down with Kaya Gangesan, a software program director at IBM, at TechXchange last fall for a Six Five demo of what was then Project Bob, getting started was as simple as downloading the IDE, entering an API key, and going. That simplicity is great for developers, but it’s also a reminder that the intelligence lived in IBM’s cloud. That’s what makes this week’s news significant. IBM has made Bob available as a self-hosted offering, and the option supports on-premises, private-cloud, sovereign-cloud, and air-gapped environments, with customers running supported models they’ve licensed locally or connecting to external model services through hybrid configurations.

IBM is extending the same option to the Bob Premium Package for Z, and that’s where the real story is. The code that most needs modernizing in banks, governments, and insurers is the code least likely to be allowed near a cloud-hosted AI service. It’s worth being precise about what “on-premises” means here, because the industry tends to blur two very different things. Nearly every AI coding tool is “locally installed” in the sense that the IDE or CLI runs on the developer’s machine, but the agent’s reasoning, the model inference, and often the repository context still travel to the vendor’s cloud. Running on-premises means the entire stack, including model serving, the agent runtime, and the governance and policy controls, sits on infrastructure the customer owns and operates. In the air-gapped case, it never touches the internet at all. That’s the difference between a tool a regulated enterprise can pilot and one it can actually put into production.

Oracle Enables Autonomous Agents with Its New Fusion Claw Offering
Oracle’s new Fusion Claw is another sign that the agent conversation is moving from assistants to harnesses built for long-running, autonomous work. Oracle describes Fusion Claw as a governed agentic execution runtime that pairs AI reasoning with deterministic enterprise computation, and it is launching with 25 Claw-powered agentic applications. The design choice I find most interesting is the division of labor. A frontier model handles reasoning, planning, and adaptation, then hands off to deterministic computation for high-volume execution, so expensive model calls are used only where intelligence is actually needed.

Governance is equally central. An “Outcome Trust Harness” enforces each organization’s policies, permissions, and approval boundaries on every run, and an “Outcome Receipt” leaves an auditable record of the decisions made and transactions executed. Customers can set each process anywhere from light assistance to fully automated execution within delegated authority. This lines up with what I wrote in last week’s newsletter about Microsoft’s new Copilot client, which includes Autopilot, an always-on agent that runs long tasks under its own identity.

Oracle’s product name may invite comparisons to OpenClaw, but the more important parallel is architectural. Vendors increasingly agree that the model is only part of the product. The harness around it, covering identity, policy, replanning, cost control, and audit, is what makes multi-hour autonomous work something an enterprise will actually allow. The difference is in the starting point. Microsoft is building a horizontal harness for knowledge workers. Oracle is building one tied to the system of record, which narrows the scope but gives its agents the business context and transactional authority that a general-purpose agent has to borrow.

Governing agents starts with knowing where they run and what each one can reach. Commvault announced a way to find agent tools on employee laptops, and CData announced a gateway that controls which company records agents can access.

I’m glad to see discovery getting attention, because a company can’t govern AI tools it doesn’t know about. Commvault announced a way for IT teams to find the agent tools employees have installed, including ones installed outside approved IT channels. The same announcement covers backing up and restoring the settings inside Anthropic’s Claude Code, Claude Cowork, and OpenClaw after a lost or replaced laptop. Those settings hold the prompt libraries and custom instructions people refine over time, and they live on the employee’s own laptop. Losing the laptop means losing work the company relies on. Whoever owns laptop backup at a company should find out what those agent tools store locally.

Once a company knows which agents are running, the next question is which company assets each one can reach. CData just announced Connect AI Gateway, a single control point between agents and company systems. The gateway includes MCP, the standard way agents connect to outside tools. It also routes each request to the most efficient model for the task. I like that the gateway ties permissions to the identity of the person asking, with the source system enforcing them down to the individual record. For example, an agent working for a salesperson would see what that salesperson could already see. That matters because an agent connected through one broad, shared login can show an employee records they could never open themselves. Companies giving agents access to customer or employee records should ask whose permissions each agent runs on.

Getting more useful work out of each expensive GPU is an old take of mine, and it continues to be top of mind. NetApp and Everpure have both announced new ways to cut the time GPUs spend waiting on storage.

I think NetApp was smart to give file lookups their own servers in its new Novus architecture. Before a GPU can read a file, storage has to look up where that file lives. A conventional array does that lookup on the same controllers that serve the data, so at GPU scale the two jobs slow each other down. Thanks to Novus, NetApp’s ONTAP storage operating system keeps the data while separate servers handle the lookups. GPU servers reach Novus through the standard NFS client built into Linux. That matters for neoclouds, the newer clouds built around renting out GPUs. A provider with thousands of servers doesn’t want to install and re-test a custom client on each one. A bank renting that capacity also retains ONTAP’s tenant-level security, and I’d expect the bank to value that above any throughput number. I went deeper in a Field Note covering all the major announcements from NetApp INSIGHT 2026.

A GPU can wait on more than file lookups. Everpure has announced PureKVA, which has its storage stage the context that a model needs into GPU memory ahead of time. Users wait less for an answer to start, and I’m happy to see storage take on that wait directly. Everpure also announced Privacy-First File Intelligence. It reports who can open each file share and which shares nobody has touched in years, without reading the files themselves. I’d want that report in hand before an AI project reads those shares. Storage already sees what a company holds and who touches it, so it’s well-positioned to answer those questions early.

CoreWeave used its Fully Connected event to tell the market that while it went public on the back of AI training, it plans to grow on inference. I spent a couple of days last week with the team in San Francisco, and I came away convinced it’s making the right moves. I also came away convinced the path is long, and that it will take significant investment, focus, and patience.

Why inference? Because raw capacity is a commodity. For the past few years, the GPU cloud business has been pretty simple: get power, get accelerators, and take orders from the model builders. It’s worked. But three customers still account for 72% of CoreWeave’s revenue, and when training growth levels off, a company that only sells GPU hours ends up in a race to the bottom on price. Inference is where AI gets used and where it drives business outcomes. That’s the durable business.

So, what is CoreWeave doing about it? I think about its strategy in two layers. Below the waterline is infrastructure, and CoreWeave does this really well. Maybe the best in the business. Mission Control, born from how its own engineers run the fleet, now gives customers a consistent operating layer across six generations of NVIDIA GPUs. Above the waterline is Forge, which brings Weights & Biases, OpenPipe, and marimo together with CoreWeave’s training and inference services. The promise? A team can take a production failure, understand it, fix it, and verify the result without leaving the environment.

Is that unique? No. AWS, Microsoft, and Google are all solving the same problem in similar ways. And honestly, CoreWeave doesn’t need to out-differentiate anyone here. It needs parity on the things enterprises care about. Its real advantages are vertical integration and focus.

The early green shoots are there. Managed inference grew from $1 million to more than $100 million in booked ARR in a matter of months. Even at its $250 million year-end target, though, that’s a small slice of an $18.5 billion to $19.5 billion run rate.

This is CoreWeave’s challenge. Caterpillar and Capital One are builders. What about everybody else? By my estimate, that’s 80% to 90% of enterprises still figuring out this “AI thing.” They want the whole product, with peer references, a known implementation cost, and someone accountable when it breaks. That’s why I think the partner network matters more than anything else CoreWeave announced. Many of these customers may never buy from CoreWeave directly. They’ll buy a solution from a partner running on CoreWeave. I’d also like to see more silicon diversity, because enterprise AI isn’t a homogeneous environment.

I expect a shakeout in the GPU cloud market. Sound familiar? I lived through the dot-com version. CoreWeave is doing what it needs to do to land on the right side of that line. Execution decides the rest. Read my full analysis here.

HPE landed its first order for the AMD Helios rack, a $1.2 billion deal with Vultr to deploy the systems across Vultr’s U.S. cloud datacenters. Each rack combines 72 AMD Instinct MI455X GPUs with EPYC “Venice” CPUs, Pensando Vulcano NICs, and ROCm software. HPE supplies the scale-up switching (six HPE Juniper switch trays per rack), direct liquid cooling, and deployment services.

Why does this matter? Because the rack-scale market is finally getting competitive. Helios has moved from an announced architecture to a real commercial deployment, and that’s good for anyone buying AI infrastructure at scale. More choice in silicon. More choice in fabric, with standards-based Ethernet and UALink over Ethernet. And more supply in a market where demand still outruns capacity.

It’s also good to see HPE deploying in the cloud. Helios lets HPE deliver compute, networking, cooling, and services as one engagement. That’s the Juniper acquisition paying off. And HPE now supplies Vultr with rack-scale systems on both NVIDIA and AMD, which is exactly how this should be played.

That said, I’d like to see HPE be more aggressive here. It has all the pieces (rack-scale integration, scale-up networking, liquid cooling expertise, and global services). The cloud and sovereign AI buildouts are happening now, and the vendors that establish deployment references early tend to keep them. Vultr should be the first of many.

And don’t overlook Vultr itself. I’ve written that I prefer the term “GPU cloud,” because most of these providers have their roots in delivering GPU capacity to the model builders. Vultr doesn’t really fit. Its roots are in cloud compute for developers, and it shows. It runs 33 cloud regions on six continents and serves hundreds of thousands of customers in 185 countries. That’s a global footprint that dwarfs most (maybe all) of its peers. It has enterprise customers. It has enterprise developers. It has a good-sized ecosystem. And it runs AMD and NVIDIA side by side. Frankly, Vultr is formidable.

What we don’t know yet is how many racks, or when they’ll come online. The real test is how quickly Helios goes from the shipping dock to operational, and whether HPE can make that repeatable.

AWS announced Built Together, a program committing more than $1 billion over five years to U.S. communities that host its datacenters. The money goes to free community college for an estimated 300,000 students, 16 new skilled-trades training centers, energy efficiency upgrades, and local grants with priorities set by each community. AWS also says it will stop using NDAs with local governments, publish its energy and water use annually, pay enough for power to keep local electricity bills from rising, and host open houses for residents.

This makes sense. A couple of weeks ago, I wrote that getting a community to say yes has quietly become the hardest part of the AI buildout. Not the chips. Not the capital. AWS itself points to more than 100 datacenter moratoriums under consideration across the country.

Here’s the part I think the industry keeps getting wrong. Community concerns come as a stack. Some have real merit. Ask anyone in PJM territory what happened to their power bill. Some are hyped. And honestly, it doesn’t matter much. Treating these communities as NIMBYs looking for a handout is a losing strategy. Treat every concern as serious and work from there.

So, does Built Together get it right? On paper, a lot of it does. It lines up well with what I argued has to hold: actual partnership, real consumption numbers published on a schedule, and committed investment with tangible benefits. Dropping NDAs is a bigger deal than it may seem. Secrecy is where suspicion starts.

But commitments are the easy part. The open houses only work if AWS shows up to listen, without a pre-canned answer for every question. And the details will matter. Will reporting be site by site, where a county can see its own water and substation load, or rolled up into a corporate figure? When a community pushes back, will a project actually change?

If the answer is yes, AWS has a much better chance of building where it wants to build. Quincy, Washington, took 20 years to become proof that partnership works. AWS has a chance to compress that timeline, and that’s an opportunity worth getting right.

I was in Barcelona last week for Microsoft FabCon Europe, where Microsoft announced a lot across the Fabric stack, from workloads and warehouses to intelligence and automation. Many of the announcements are about making Fabric easier for a company to start with.

Companies that have run Power BI for a few years already have something that agents need. Their semantic models hold business definitions that people keep fixing whenever a report’s numbers look wrong. Microsoft just made Fabric IQ generally available in Copilot, so Copilot now answers business questions from those models, metrics, and definitions. It’s on by default for Fabric and Power BI customers, and existing access controls and governance stay in place. I’m a fan of context that comes from work the business already maintains, because those definitions stay current as a side effect. The flip side is that Copilot now inherits whatever sits in each model — stale ones included. I’d give the important models an owner, clear descriptions, and a timely review process.

Then there’s the billing change that got the loudest cheer in the keynote room, and I understand why. Teams have to size capacity before they know how much they’ll use, and that’s hard to get right. Microsoft announced on-demand billing, where a workload is charged only for the compute it uses. The company also announced Fabric zero-provisioned, which lets a team start on Fabric and OneLake with no upfront capacity commitment. I’d expect usage billing to suit agents, since it’s hard to predict how many jobs an agent will start. Steady workloads can still sit on reserved capacity, so a team doesn’t have to switch everything. And teams on the older Azure Data Factory service may see little reason to move to Fabric when that service already works. Zero-provisioned gives those teams a low-risk way to try Fabric.

The Witcher 3: Wild Hunt — Remastered is getting a free update that adds path tracing, DLSS 4.5, and other enhancements, and is playable now through GeForce NOW. This is a big deal because it is sure to significantly improve the image quality and performance of a very popular game, and many people are excited to play it with better graphics. Hopefully NVIDIA will be able to give it a DLSS 5 update as well, but that seems a little more controversial and might require more time with the developer.

Cognex agreed to acquire RealSense for about $500 million in cash, adding 3-D depth sensing and robot perception to its machine-vision business. The company expects to close the deal in Q4 2026. RealSense began inside Intel in 2014, became independent in 2025, and expects $80 million to $90 million in revenue this year. Cognex targets fixed-arm robots and autonomous mobile robots, and plans humanoids and quadrupeds as longer-term opportunities. RealSense’s D555 camera uses Ethernet for power and data, runs vision processing on the camera itself, and publishes depth data directly into ROS 2 robot software over the network, with no USB-connected PC in between.

My take: Industrial vision is expanding from fixed-position inspection to mobile 3-D perception. RealSense bridges the gap between Cognex’s high-end factory inspection and high-volume mobile robotics: autonomous mobile robots and cobots. It also brings more than a million cameras shipped and about 20,000 active developers. Cognex says it may run its own AI models on RealSense hardware, which would bring its inspection software to moving robots.

The Wireless Broadband Alliance and the FIDO Alliance published trial results for zero-touch Wi-Fi onboarding of IoT devices. The approach combines OpenRoaming/WBA’s system for joining Wi-Fi networks automatically with FIDO Device Onboard (FDO), which lets a device ship with credentials and transfer ownership to its buyer at installation. In the trials, devices connected securely at first power-on, began ownership transfer, and received operational credentials without manual intervention. This solution does not currently address air-gapped networks and constrained devices that cannot run FDO protocols. The report explores an interesting proxy model for constrained devices in which a helper device runs the onboarding protocols.

My take: The industry has been slow to warm up to FDO, largely because the concept’s technical details aren’t easy to grasp at a glance. However, if you need to onboard thousands of devices without physically configuring each one, it’s worth your time to dig into the technology. The constrained-device gap is where most sensors live, and a helper device that onboards its neighbors sounds like a good solution. I’m happy to see FDO get more well-deserved attention. (In 2018, I worked with Intel on the technology that became FDO.)

The big three U.S. carriers have made their satellite joint venture official and appointed an interim CEO, Paul Roth. This joint venture was originally floated back in May of this year (I wrote about it at the time) but obviously took time to organize and clear regulatory hurdles. The carriers are framing it as a joint solution to dead zones, allowing them to pool spectrum and deliver better coverage through direct-to-device (D2D) satellite links. The companies also said that this JV will not preclude any existing satellite partnerships that any of the carriers already have in place. However, this whole JV is clearly a hedge against SpaceX’s Starlink efforts and its attempt to launch a standalone service. SpaceX seems to feel threatened, judging by its executives’ comments on the JV, including calling it a “cartel.” If anything, I think this JV will help the operators simplify their offerings to consumers, even if that doesn’t mean unifying under a single service.

Panasonic announced the next phase of its TOUGHBOOK lineup with the TOUGHBOOK 34 and G3 convertible PCs. These are 12.4- and 10.1-inch 2-in-1 convertible laptops that bring some of the latest tech to the ruggedized space. This means using Intel’s Core Ultra series of processors while still offering modularity and cellular connectivity. The TOUGHBOOK 34 is certified for Red Hat Device Edge, meaning the notebook can double as a local AI compute solution for smaller devices without complex connectivity or processing. This increases the value of the PC itself as well as its importance to IT when trying to solve for local computing issues. (For more on the TOUGHBOOK series, see this Research Note from May.)

AMD’s acquisition of World Labs for $8.2 billion surprised many because most people don’t consider AMD that much of a physical AI company. World Labs is all about physical AI, with automotive and robotics at the top of its list. I believe this move indicates that AMD is getting more serious about physical AI, and telegraphs to both its customers and the industry at large that it is heavily pursuing this area. I’m not entirely sure this was the right approach, and I do think AMD will need to start producing customers in physical AI to justify this purchase.

Synopys signed a $1 billion multi-year silicon IP agreement with Amazon. This deal also has a license-plus-royalty model, so there’s significant upside for Synopsys. Unsurprisingly, most of this announcement centers on Amazon’s own silicon, Trainium and Graviton, used for training and inference. This collaboration also includes engineering software and planned optimizations for Amazon silicon, potentially enabling Amazon to run this software on its own silicon in the near future.