A new Microsoft Copilot client and runtime could accelerate the citizen developer movement
Last week, Microsoft announced a major rework of Copilot. It folds Chat, Cowork, a new Code mode for building apps in natural language, and Autopilot, its long-running autonomous agent, into a single desktop client. Alongside it, Microsoft launched Copilot Managed Runtime in public preview. The runtime is a governed hosting environment inside the Microsoft 365 tenant, and it applies Entra identity, organizational policies, source control, and a central admin inventory to apps built with Copilot, Copilot Studio, GitHub Copilot, or partner tools such as Lovable. My take is that this pairing removes many of the barriers that have held back citizen development for decades. Demand from knowledge workers is already there, and Microsoft has separated where apps are built from where they are run and governed, which no monolithic low-code platform ever did. It is not a finished replacement for low-code platforms yet. Managing dormant apps, portability, and cost predictability under usage-based pricing are all still open questions. For now, I see it as a new answer to an old problem, and one that IT leaders should pay close attention to. You can read more in my latest Field Note.

Yet again, frontier companies disappoint me…
Thanks to Evan Schuman for reaching out and publishing this selection of analyst viewpoints on OpenAI’s recent third-party testing principles. Frankly, OpenAI’s statement lacked real teeth in terms of escalations and resolution if and when evaluators do find an issue. That lack of rigor, combined with what we have seen in the past couple of weeks (“Let’s pace! But not until I get a new model out”), is likely to be met with an eyeroll. But where some will see market posturing, I see the beginning of what could be a long negotiation with competitors, partners, and regulatory bodies.

QAD continues the trend towards specialized agents
At QAD | Redzone’s Champions of Manufacturing 2026 event in Chicago, I got the clearest look yet at a bet the company has been making for about a year — that mid-market manufacturers, largely ignored by enterprise AI vendors building for big IT shops, are the right place to plant an agentic flag. The product cadence backs that up, with Champion Assist, new Accounts Payable, Trade Compliance, and Lynx Champions, and seven new frontline personas layered on top of a Redzone footprint that QAD says spans more than 2,000 plants. Lynx stands out because it attacks the core reason mid-market firms fall behind on ERP (re-platforming is too costly and risky) by automating discovery and scoring a modernization path. The bigger news may be Manufacturing Intelligence, a cross-process layer QAD plans to ship in spring 2027 that leans on NVIDIA vision models for shop-floor inspection and paper-heavy workflows; it’s the first real sign that QAD’s AI ambitions extend beyond back-office agents to the plant floor.

My take: QAD’s manufacturing-only focus gives it a genuine structural advantage over horizontal players like SAP, Oracle, and Microsoft right now, but focus without connective tissue becomes merely a walled garden. QAD needs to prove its Boomi-based integration and “works with any ERP” claims with named partners and live customers outside its own stack, or that advantage could erode quickly.

For more on this, see my latest Analyst Insight.

A new Microsoft Copilot client and runtime could accelerate the citizen developer movement
Last week, Microsoft announced a major rework of Copilot. It folds Chat, Cowork, a new Code mode for building apps in natural language, and Autopilot, its long-running autonomous agent, into a single desktop client. Alongside it, Microsoft launched Copilot Managed Runtime in public preview. The runtime is a governed hosting environment inside the Microsoft 365 tenant, and it applies Entra identity, organizational policies, source control, and a central admin inventory to apps built with Copilot, Copilot Studio, GitHub Copilot, or partner tools such as Lovable. My take is that this pairing removes many of the barriers that have held back citizen development for decades. Demand from knowledge workers is already there, and Microsoft has separated where apps are built from where they are run and governed, which no monolithic low-code platform ever did. It is not a finished replacement for low-code platforms yet. Managing dormant apps, portability, and cost predictability under usage-based pricing are all still open questions. For now, I see it as a new answer to an old problem, and one that IT leaders should pay close attention to. You can read more in my latest Field Note.

Yet again, frontier companies disappoint me…
Thanks to Evan Schuman for reaching out and publishing this selection of analyst viewpoints on OpenAI’s recent third-party testing principles. Frankly, OpenAI’s statement lacked real teeth in terms of escalations and resolution if and when evaluators do find an issue. That lack of rigor, combined with what we have seen in the past couple of weeks (“Let’s pace! But not until I get a new model out”), is likely to be met with an eyeroll. But where some will see market posturing, I see the beginning of what could be a long negotiation with competitors, partners, and regulatory bodies.

QAD continues the trend towards specialized agents
At QAD | Redzone’s Champions of Manufacturing 2026 event in Chicago, I got the clearest look yet at a bet the company has been making for about a year — that mid-market manufacturers, largely ignored by enterprise AI vendors building for big IT shops, are the right place to plant an agentic flag. The product cadence backs that up, with Champion Assist, new Accounts Payable, Trade Compliance, and Lynx Champions, and seven new frontline personas layered on top of a Redzone footprint that QAD says spans more than 2,000 plants. Lynx stands out because it attacks the core reason mid-market firms fall behind on ERP (re-platforming is too costly and risky) by automating discovery and scoring a modernization path. The bigger news may be Manufacturing Intelligence, a cross-process layer QAD plans to ship in spring 2027 that leans on NVIDIA vision models for shop-floor inspection and paper-heavy workflows; it’s the first real sign that QAD’s AI ambitions extend beyond back-office agents to the plant floor.

My take: QAD’s manufacturing-only focus gives it a genuine structural advantage over horizontal players like SAP, Oracle, and Microsoft right now, but focus without connective tissue becomes merely a walled garden. QAD needs to prove its Boomi-based integration and “works with any ERP” claims with named partners and live customers outside its own stack, or that advantage could erode quickly.

For more on this, see my latest Analyst Insight.

Most governance falls apart the moment data leaves the system that governs it. That moment can be as big as a bank sending customer records to someone else’s model, or as small as an analyst exporting a table to Excel. VAST just took on this problem from the model side, and Databricks and ThoughtSpot took on the spreadsheet angle.

Regulated companies and model builders have been stuck in a standoff. A bank can’t send its customer records to an outside model. The company that built the model won’t hand over its weights — i.e., the trained core that makes it worth paying for. To resolve this, VAST just launched VAST DataEnclave, a confidential AI capability for the VAST DataEngine, so it switches on inside a platform VAST customers already run. The model works on the data in a protected area of the processor that even the hardware operator can’t read. Keys to the data are released only after the hardware proves it’s running the approved setup. I really like that every check and key release lands in a queryable record, since auditors eventually ask what ran. Check out my deep dive on DataEnclave.

Historically, and in my experience, far more data leaves through spreadsheet exports than through model leaks. Every export creates a copy with no permissions or audit trail. To address this, Databricks just bought Row Zero, which makes a cloud spreadsheet that stays connected to live, governed data. Permissions follow each user, and every write-back gets tracked. Meanwhile, ThoughtSpot launched Spreadsheets on Live Data, which carry its data models’ security rules into the grid and write results back to the warehouse. I love the idea of getting sensitive data off laptops without asking anyone to give up their spreadsheet. Letting people write changes back is a nice touch too, and I’m curious how each team decides which edits get in and how they’re checked.

AI keeps landing in parts of the stack that were built for a smaller job. For example, file systems were sized for people opening files, and monitoring was built to tell you whether an app is up. NetApp and AWS are each rebuilding one of those parts.

Every file system keeps an index of its files, with names, locations, and permissions. That index is called metadata. At GPU-cluster scale, it tends to choke well before the disks do, because thousands of GPUs are requesting millions of small files. NetApp just announced its intent to acquire PEAK:AIO, a U.K. company whose parallel file system already runs at Los Alamos National Laboratory and Carnegie Mellon University. The plan would give metadata its own tier inside ONTAP, so it can scale separately from the data. I like that NetApp is aiming straight at the bottleneck. This is NetApp’s third deal since July. In the previous two deals, DataPelago brought processing to where the data sits, and JetStream opened a path into VMware estates. PEAK:AIO would change how ONTAP scales for AI clouds running large GPU clusters.

The agents running on those clusters fail in ways that traditional monitoring wasn’t built to catch. An agent can look perfectly healthy in terms of speed and errors while giving customers wrong answers for days. AWS just launched Amazon CloudWatch Omni, which traces the model calls, tool calls, and reasoning steps an agent takes and scores its answers against live traffic. A team can now watch answer quality as closely as it already watches uptime. I’m a fan of AWS building Omni on OpenTelemetry and supporting the agent frameworks that teams already use. As I told InfoWorld, those scores are only as good as a company’s definition of a good answer. Plenty of companies haven’t written that definition down yet.

These days, more of the work on a data platform gets done by agents and coding assistants, with a person reviewing the result. So the platforms are competing on what they hand those tools, from business definitions to the right data for each step.

Every time an agent works out what “revenue” means from scratch, for example, somebody pays for model calls. Teradata’s Tera Context Engine keeps a company’s business definitions in a governed layer, so agents look them up and skip the reasoning. Keeping definitions in a lookup helps with cost control, which is smart, since agents make consumption bills hard to predict. The Tera Harness picks the skills, tools, data, and model for each step — a choice the earlier Tera left to a person. Teradata calls the context engine vendor-neutral, so it reaches even data that Teradata doesn’t manage. Human approvals sit inside the agent loop, which I’m glad to see, since one bad pick by the harness could otherwise carry through every later step. And because a paused job doesn’t burn compute while it waits for sign-off, those approvals don’t add to the bill.

Plenty of developers I talk to now build in Claude Code or Cursor, and they’d rather not leave. At its annual Succeed conference, Dataiku launched Headless as part of its expanded Cobuild agentic interface so those developers can build pipelines and agents from their preferred coding tool. I’ve been pretty clear from the beginning that meeting developers where they already work is the right play. The AI Catalog in Cobuild now covers semantic models too, so anything those developers build shares the business definitions everyone else in Dataiku uses. Then there’s Agent Management, which inventories agents across platforms from AWS, Salesforce, and Snowflake, including agents Dataiku didn’t build.

Meta announced what it is calling VR Glasses, powered by a best-in-class Snapdragon Reality Elite chip. The glasses weigh only 100g and are connected via a fiber-optic cable to a compute puck that removes the weight and heat from the glasses themselves. This tethered approach is a way to make VR more comfortable for longer wear, which seems to be Meta’s goal here. It seems to want people to spend more time in VR, which it believes will drive more usage and enhance people’s engagement. While Meta is still talking about using the new glasses for gaming, that does seem to take a back seat to entertainment and productivity, which tend to involve longer sessions. That said, Meta’s success in VR gaming so far has come at price points of $299 to $499; by contrast, these glasses are $1,299, which I believe aims at an inherently different audience. Meta VR Glasses also allow external devices to connect via the compute puck, so you should be able to easily connect your laptop either wirelessly or via a wired connection. The oddest part of this announcement is that Meta is waiting until spring 2027 to make the VR Glasses available, which is roughly half a year away — a long gap after the announcement.

Samsara made its MCP server generally available on September 22. Customers can add the hosted endpoint to Claude, ChatGPT, Copilot, Cursor, or any AI tool that supports the Model Context Protocol, sign in with their Samsara login, and their agents get permissioned read-only access to Samsara data. The agents can query vehicles, drivers, safety events, hours of service, and more through more than 40 Samsara tools. For more complex work, the agents can send a plain-language request to the Samsara Assistant, which does the reasoning inside Samsara, backed by Samsara’s purpose-built agents. Every request carries the user’s Samsara permissions, with no shared API keys, and every call is attributable to a named user. MCP access is free today, and Samsara says future pricing will be based on usage. Samsara calls itself the system of record for physical operations and says it is building toward a “headless Samsara” that makes the tools and intelligence behind the Samsara dashboard available on every surface.

My take: At the Samsara Beyond event in June, I asked about providing direct access to MCP servers, and this announcement is the answer. Agents supplement, and may eventually replace, the dashboards and control panels that traditionally interact with Samsara’s massive operational database. Agentic AI is rapidly becoming the new UI for physical systems across many industries, and Samsara is looking to make this transition easier.

Qualcomm has agreed to acquire PickNik, the Boulder, Colorado, company that stewards MoveIt, one of the most widely adopted open-source manipulation frameworks in the Robot Operating System (ROS) ecosystem. The companies demonstrated MoveIt Pro running on the Arduino VENTUNO Q board and Dragonwing robotics platforms at ROSCon in Toronto. Qualcomm plans closer integration of MoveIt with Dragonwing and Arduino hardware, and says that MoveIt 1 and MoveIt 2 will remain open-source and continue supporting third-party hardware and ROS. Terms were not disclosed. MoveIt reaches beyond PickNik’s own customers: NVIDIA’s cuMotion planner ships as a MoveIt 2 plugin.

My take: Vision-language-action (VLA) models map vision and language instructions to robot actions. MoveIt turns those decisions into safe motion through kinematics, collision avoidance, and motion planning. In simple terms, PickNik puts the A in VLA. The deal also extends Qualcomm’s platform strategy: Foundries.io for productization, Edge Impulse and Arduino for developers, Modular for AI software, and now PickNik for manipulation. Qualcomm is building a comprehensive robotics platform, and the open-source commitment is in the spotlight as developers evaluate these tools.

The Thread Group announced on September 21 that its board of directors approved development of Thread on a sub-GHz radio, targeting commercial and industrial applications. The technology will carry a distinct brand. The group did not announce the brand name, specific frequencies, or a timeline. Sub-GHz signals reach farther and pass through walls and floors better than 2.4 GHz radios, and the band is currently less crowded. The trade-offs are lower data rates and region-specific band plans: 868 MHz in Europe and 915 MHz in North America. IEEE 802.15.4, the radio standard under Thread, has defined sub-GHz radios since 2003. The 802.15.4g amendment in 2012 added the sub-GHz radios used by Wi-SUN for utility meshes. Z-Wave has always operated in the sub-GHz range, and the Connectivity Standards Alliance opened certification for Suzi, Zigbee’s sub-GHz extension, this September.

Thread has resisted sub-GHz for years. In 2023, the Thread Group formally announced that its 2.4 GHz mesh remained the best fit, even though commercial and industrial sub-GHz networks have been around for decades. Here’s the key question. It’s great that Thread is going multi-PHY, but can Thread 2.4 GHz and sub-GHz devices coexist on the same network? Today, Wi-Fi can extend Thread networks, so I assume the answer is “yes,” provided there’s a border router or a full Thread device with both radios, and provided the specifications enable coexistence and routing. However, public statements from the Thread Group haven’t specifically addressed this issue.

Disclosure: I’m keynoting the October Thread Group member meeting.

AST SpaceMobile announced a $1 billion joint venture with Rakuten to deliver direct-to-device (D2D) service in Japan. The Japanese government is backing the joint venture with $926 million in subsidies. The 700 MHz spectrum could enable nationwide coverage by fiscal 2027. This gives AST SpaceMobile yet another large carrier deal beyond its AT&T and Vodafone deals, which should help the company reach the global scale it needs for profitability. AST SpaceMobile still faces stiff competition from SpaceX Starlink and Amazon Leo, but its novel approach uses much larger satellites.

At the Snapdragon Summit in Hawaii, Qualcomm finally announced support for Linux, starting with Debian and Ubuntu. This shift should be meaningful for the PC market because AMD, Intel, and NVIDIA already support Linux. Linux continues to be a big growth driver in the PC market, and I see it as something power users have preferred over Windows in recent years. While Qualcomm hasn’t said anything about SteamOS, introducing support for it could be a natural progression, especially since SteamOS already runs on Qualcomm’s hardware in the Steam Frame. I can’t imagine Microsoft is happy about this, considering the close relationship it has with Qualcomm, but it seems clear that a large segment of PC users now prefer Linux over Windows.

Also at the Snapdragon Summit, Microsoft launched two new Surface PCs and a new Surface Haptic Mouse. The new PCs are a second-generation Surface Pro 12 and Surface Laptop 13, both with Qualcomm’s Snapdragon X2 Plus, its mid-range chip. The Surface Pro 12 also comes with 5G connectivity, which makes it a highly portable, well-equipped machine for its small form factor. Microsoft also updated some apps, including an enhanced Canvas app with AI-assisted inking.

1. Non-Abelian anyons demonstrate universal quantum gates. Researchers used 54 trapped-ion qubits on Quantinuum’s H2 processor to create a state with non-Abelian anyons. The team combined braiding the anyons with fusion measurements to demonstrate a set of operations that can, in principle, support universal quantum computing. Braiding alone was not enough. Encoding information across the anyons may help protect it from local noise, but the experiment did not establish that this approach can yet deliver large-scale, fault-tolerant computing.
   
   
2. Stanford physicists made the first direct, real-time observation of a quantum jump in sound. They repeatedly measured a tiny mechanical resonator and tracked a sudden change from a state containing one phonon — a unit of vibration — to a state containing none. Scientists had previously observed quantum jumps in other systems and had found evidence of phonon jumps, but had not directly tracked an individual phonon jump in real time. The method could eventually help with quantum devices and sensitive sensors; those applications were not demonstrated in this experiment.
   
   
3. IonQ researchers tested an error-correction decoder on a single conventional CPU. They fed it simulated error data from proposed trapped-ion computations, including a benchmark representing up to 408 logical qubits. In one favorable test, decoding added 0.02% to the computation’s running time. That number measures added time, not the decoder’s latency, and other tested conditions produced more delay. The work suggests that a conventional CPU could handle decoding for this particular architecture under the assumptions tested. It does not show a live 408-logical-qubit computer performing error correction or prove that specialized hardware will never be needed.

Qualcomm unveiled its dual-flagship strategy at its Snapdragon Summit event. With a new Snapdragon Sound Elite chip for wearables, Qualcomm is building an entire lineup of AI-capable chips. While the Snapdragon 8 Elite Gen 6 and Elite Extreme Gen 6 continue to push Qualcomm’s position against Apple, the company also talked a lot about agentic capabilities. It positioned these chips to power future agentic devices, and given the breadth of what Qualcomm offers today, that makes a lot of sense. Qualcomm has not only the CPU, GPU, NPU, and connectivity IP, but also the platforms and scale to allow virtually anyone to build an AI device. For example, all of Meta’s new AI smart glasses, VR glasses, and Muse Charm personal AI device run on some form of Qualcomm chip.