The App Store is evolving at a breakneck pace not seen since its earliest days. We saw the first glimmers of what was to come early in 2025, but it wasn’t until late last year that the tsunami of apps developed with the help of AI agents really took hold.
Since then, veteran developers are releasing new apps and updating existing ones faster than ever, and new developers are releasing their first apps in droves. Today, supply is dramatically up, demand is flat, and quality is seemingly simultaneously up and down, depending on where you look.
How these forces play out long-term is anyone’s guess, but it’s worth examining because, just as Federico’s link to Bryan Irace’s post about agentic coding tools foreshadowed the rise of tools like Codex and Claude Code, today’s trends are sparks that have the potential to become tomorrow’s App Store wildfires or simply fizzle out.
Utility apps are on the front lines of this change. The TechCrunch story I linked in April picked up on this trend:
Another interesting tidbit from Appfigures is that the Utilities app category moved up the top five chart.
If anything, the trend has accelerated in the months since.
It’s not surprising at all that utility apps have taken off. They’ve been a staple of new developers since long before agents came along. That’s because many are UI wrappers around command line tools. That isn’t a knock against the developers of those apps; most users don’t want to open Terminal to convert a video or audio file to another format using ffmpeg, for example. By building a great UI around command line tools, developers have made them far easier to use.
However, the relative simplicity and narrow scope of many utility apps have made them a natural fit for AI agents, too. That’s why the App Store (and my inbox) is deluged with Mac menu bar and single-screen iPhone and iPad utility apps.
On the one hand, the abundance of utilities has been great for users. More choice means you’re more likely to find the app that perfectly fits your needs.
On the other hand, though, I don’t think what’s happening in the category is sustainable and expect to see it dramatically shift again in the coming months. As we’ve seen from reporting by The New York Times, app supply is outstripping demand by orders of magnitude, which will drive down prices. That alone is likely to cause the utility app market to shrink. But there’s more to it than that.
Remember, the demand shifts that The New York Times reported, based on Sensor Tower numbers, are for the entire App Store, where download numbers have grown 2-3% this year and last. I expect utility downloads to actually shrink in the coming months – again, because of agents. Utilities, especially simple ones, are exactly the sorts of apps that are becoming trivially easy for power users – the very users these utilities are made for – to create themselves. From frontier labs‘ model improvements to a growing number of app-building tools from those same labs and third parties, it’s never been easier to build a web or native app yourself.
And although I agree with people, like Nilay Patel, who say the notion that everyone will build their own software is overblown, the utility market is different. Your average person is not downloading apps to adjust the frame rate and file format of a video or downloading YouTube videos to watch later. But those are exactly the sort of things that people who are using agents are doing with them, whether they’re having an agent do those things directly via a command line tool in an app like Codex or Claude Code or building an app to do the same thing. That’s going to put even more pressure on the category.
That said, I think there will always be a cohort of users who would rather pay for an app than build it themselves, so I’m not predicting the demise of utilities in general – just the end of today’s frothy market. I also think there remains a place for simple utilities that solve hard problems with clever solutions. Not every utility is a UI wrapper for a terminal command. Plenty of apps feature original solutions or thoughtful and unconventional remixes of disparate tools.
Utilities aren’t going away, but just like other App Store deluges, the trend will recede. It’s just that this time, it’s likely to flip faster than usual.
OpenAI family of devices plans are no longer limited to an experimental gadget hidden inside a design studio. OpenAI President Greg Brockman has confirmed that the company’s first consumer device will begin a broader series intended to reconsider how people interact with artificial intelligence.
The confirmation gives OpenAI’s hardware strategy a clearer shape without revealing the first product itself. Brockman said the company is not ready to discuss the device in detail, but described it as the first in a series that will rethink what devices should be when AI can understand language, observe context and complete tasks.
That ambition extends well beyond placing the ChatGPT app on another piece of hardware. OpenAI wants to reduce the menus, buttons, model selectors and separate applications that currently stand between a person and an AI assistant.
For Apple, Google and other established platform owners, the implication is significant. ChatGPT currently reaches consumers through devices and operating systems controlled by other companies. An OpenAI-designed product would give the company direct control over the microphones, cameras, sensors, interface and account experience through which its models interact with people. 56000
OpenAI’s move into consumer hardware accelerated in 2025 when it acquired io Products in a transaction valued at approximately $6.5 billion.
The startup had been founded by former Apple design chief Jony Ive alongside Scott Cannon, Evans Hankey and Tang Tan. Its team included hardware engineers, designers, manufacturing specialists and product developers, many of whom had spent years working together at Apple.
OpenAI and Ive had already been collaborating before the acquisition. Their early ideas developed into physical designs, convincing the partners that they needed a dedicated company capable of engineering and manufacturing complete products rather than another software prototype.
The io team officially merged with OpenAI in July 2025. Ive and his design collective LoveFrom remained independent while assuming extensive design and creative responsibilities across OpenAI.
The structure combines OpenAI’s models and software research with a group experienced in turning new interface concepts into mass-produced consumer products. Ive’s previous work helped shape the iMac, iPod, iPhone, iPad, Apple Watch and other devices that established Apple’s modern design identity.
OpenAI has continued recruiting hardware expertise from Apple. Paul Meade, who led hardware development for Apple Vision Pro and worked on the company’s smart-glasses efforts, reportedly joined OpenAI’s hardware organization in 2026.
OpenAI also placed product-development veteran Peter Welinder in a leadership role connected to the io division. The resulting organization is not structured like a small accessory project. It is being assembled to develop products, create manufacturing relationships and support more than one form factor.
OpenAI has not publicly confirmed the name, appearance, price or specifications of its first device. It has also not presented the product as a smartphone, pair of glasses, wearable pin or conventional smart speaker.
Recent reporting describes a portable, screen-free device designed primarily for use inside the home. The reported product resembles a compact speaker but includes a rechargeable battery, allowing it to move between rooms rather than remain attached to one location.
A camera and additional sensors would help the device understand its surroundings. Mechanical elements could allow parts of it to move, creating a more expressive physical presence than a stationary speaker waiting for a wake word.
The reported design would rely heavily on OpenAI’s latest voice technology. Instead of requiring a person to wait for the assistant to finish speaking, the system could listen and respond more naturally during overlapping conversation. Access to personal information, including messages or email when authorized, could give it enough context to handle requests rather than merely answer questions.
These details remain reported plans rather than announced specifications. OpenAI could change the hardware, features or release schedule before introducing the product.
The company is reportedly targeting a presentation during 2026 and a commercial release in 2027. Fortune separately reported that the initial design has been finalized, although development of the larger product family continues.
ChatGPT already operates across iPhone, iPad, Mac, Android and Windows devices. That wide distribution gives OpenAI access to an enormous audience without requiring it to manufacture screens, batteries or processors.
It also leaves crucial parts of the experience under the control of Apple, Google and Microsoft.
An operating system determines when ChatGPT can use a microphone, whether it can run continuously, how notifications appear and which applications it can control. Platform owners can limit background access, change default-assistant rules or promote their own AI systems more prominently.
Those restrictions matter more as ChatGPT develops from a chatbot into an agent. Answering a question inside an app requires relatively little control over the device. Completing a task across messages, calendars, files, cameras and connected appliances requires deeper access.
OpenAI hardware could be designed around those capabilities from the beginning. The assistant would not need to behave like one application among dozens because it could become the primary interface.
Brockman has described the current ChatGPT interface as too complicated, particularly as new models and tools add more controls. OpenAI’s objective is to make much of that interface disappear.
A successful device could allow someone to describe an outcome instead of opening several applications and completing each step manually. The user might ask the assistant to reorganize a day after a delayed flight, contact the relevant people and adjust reservations. The interface would be the request and the resulting confirmation rather than a sequence of taps.
Calling the project a family of devices changes the strategy. One experimental product would need to prove its value independently. A coordinated group could distribute different sensors and functions across the places where people already live and work.
A home device could provide shared access in kitchens and living spaces. A personal wearable could preserve private conversations while moving outside. A future mobile product could handle visual information, navigation and communication without reproducing the application grid of a smartphone.
OpenAI is reportedly developing approximately five hardware concepts, with the screen-free home companion receiving priority. The company has also explored wearables and home robotics, although none of those later products has been announced.
The devices would become more useful if they shared the same ChatGPT identity, memory controls and task history. A request started at home could continue through another product without requiring the user to repeat the people involved, the desired result or the decisions already made.
That continuity is one of the strongest advantages Apple currently receives from controlling a complete ecosystem. An iPhone, Mac, Apple Watch, AirPods and HomePod can exchange information because Apple designs the devices, operating systems and cloud services together.
OpenAI appears to be pursuing a different center of gravity. Instead of making the operating system coordinate applications, it could make the AI assistant coordinate specialized devices.
OpenAI’s hardware work does not mean the company intends to reserve ChatGPT for its own products.
Brockman has said OpenAI is interested in deeper integrations with smart glasses, wearables and other devices through its application programming interfaces. The company wants ChatGPT to become an interface available across hardware categories, including products manufactured by third parties.
That creates a two-track strategy. OpenAI can use its own devices to demonstrate the experience it considers ideal while allowing other manufacturers to connect their cameras, microphones and sensors to ChatGPT.
Third-party distribution would help OpenAI reach consumers who do not want another dedicated device. It would also reduce the risk of depending entirely on the success of one hardware launch.
The difficult part will be keeping those experiences consistent. A wearable may have limited battery capacity, a camera may require visible privacy controls, and a speaker may depend on cloud processing when a connection is available. OpenAI will need to define what its assistant can do reliably across products with very different technical limits.
The first generation of dedicated AI gadgets established a warning for every company entering this market.
Humane’s AI Pin and the Rabbit R1 attracted attention by promising an alternative to application-based computing. Their launches also exposed problems involving response delays, battery life, incomplete features and tasks that remained easier on a smartphone.
Those products demonstrated that access to a language model does not automatically justify a new object. Consumers need to understand why they should carry, charge and maintain another device when ChatGPT, Gemini and other assistants already operate on phones.
OpenAI has several advantages those earlier companies lacked. It controls one of the most widely used AI services, operates the models behind the interface and has the resources to improve the software after launch. The io acquisition gives it experienced design and manufacturing leadership rather than requiring it to assemble a hardware organization gradually.
The standard will consequently be higher. A well-designed enclosure and natural voice will not be enough if the device cannot complete dependable tasks more efficiently than an iPhone or Android phone.
OpenAI’s expansion beyond Apple devices has also produced a direct legal confrontation with Apple.
Apple filed a federal lawsuit in July 2026 accusing OpenAI, io and former Apple employees of misappropriating confidential hardware information. The complaint alleges that trade secrets connected to unreleased products, engineering processes and suppliers were taken for the benefit of OpenAI’s hardware business.
OpenAI has denied having any interest in Apple’s trade secrets and said it was unaware of evidence giving the complaint merit. The allegations have not been decided by a court.
Apple is seeking restrictions on the use of its confidential information, along with the return of materials and financial damages. The case could require OpenAI to separate employees from parts of the project, replace suppliers or repeat engineering work if disputed information is found inside the development process.
It does not automatically establish that OpenAI’s reported product copies an Apple device. The first OpenAI design is described as substantially different from Apple’s current hardware, but the legal dispute could still affect when it can be manufactured or sold.
The most important details at OpenAI’s eventual hardware presentation may have little to do with the object’s shape.
A camera-equipped device that learns from activity inside a home will need unmistakable recording indicators, understandable data-retention controls and a simple method for temporarily disabling its sensors. Owners will need to know which information remains on the device, what reaches OpenAI’s servers and which memories can be reviewed or deleted.
OpenAI will also have to explain whether the product requires a paid ChatGPT subscription, how much it can do without an internet connection and whether multiple members of a household receive separate identities. Those choices will determine whether it behaves like a personal assistant, a shared appliance or a new subscription terminal.
Support for messages, calendars, music and connected appliances will require cooperation from outside companies. A device cannot replace frequent phone interactions if it remains unable to enter the services where those interactions occur.
The first model will therefore establish more than OpenAI’s industrial-design language. It will set the privacy, account, subscription and developer rules that every later member of the OpenAI family of devices inherits.
The post OpenAI Family of Devices Takes ChatGPT Beyond Apple appeared first on AppleMagazine.
The notion that scammers can use AI to sharpen their deceptions, polish their language, and lubricate their banter with victims is now a reality for anyone fighting the fraud operations that steal tens of billions of dollars a year worldwide. But can AI fully replace a human scammer, autonomously building the web of deception leading up to the fake investment that defrauds the mark? One study's experiment suggests that it can—and may even be able to carry out the majority of that long con more effectively than humans.
Researchers from four universities—Amrita Vishwa Vidyapeetham in India, Foscari University of Venice, the University of Melbourne, and Ben Gurion University of the Negev—carried out a broad study on the use and potential of generative AI chatbots in the growing scam industry centered around a form of fraud known as “pig butchering,” text-based romance scams that eventually shift to fake crypto investments that steal as much as six-figure sums from victims. In their study, the researchers pitted AI chatbots directly against humans in a simulation of the scamming process—or more specifically, the long, trust-building conversations that eventually lead up to soliciting a fake investment from the scam’s target.
They found that for the relationship-establishing stages of the scam—the stage that in real-world scams typically represents the longest part of the interactions with the victim, often stretching to months—an AI chatbot performed remarkably effectively, successfully impersonating a human and by some measures outperforming the real human “scammers” in their experiment.
© Nansan Houn/Getty
Kalley Huang, writing for The New York Times about apps written with the help of AI agents that are flooding the App Store:
But as with many things A.I., just because something is easy to build doesn’t mean people will use it. It is not clear if vibecoding is breathing new life into the App Store or just cluttering it.
Last year, the number of new apps released in the App Store grew 30 percent to about 600,000, according to estimates by Sensor Tower, an app analytics firm. In the first half of this year, new apps doubled to about 560,000.
Now, Sensor Tower metrics should always be taken with a grain of salt. They don’t have direct access to Apple’s sales data, so they’re extrapolating from incomplete data that they collect themselves. That said, I think it’s fair to take their numbers as broadly representative of sales trends over time, and the story they tell, as reported by Huang, is interesting.
Based on Sensor Tower’s numbers, the total number of App Store releases peaked in 2016 at 890,000, hit a low of 420,000 in 2022, and this year, is on pace to pass 2016’s peak by a healthy margin. But app releases don’t equate to downloads, let alone sales. According to Huang’s reporting, downloads increased 3% in 2025 and 2% in the first half of 2026, far lower than the growth of new releases.
All of this tracks closely with what we’ve seen at MacStories. One subtle trend I’d add is that whereas for years most developers contacted us before they released an app, a lot of new developers are doing so after their apps are on the App Store. I suspect what I’m seeing is a new generation of developers learning the hard discoverability lessons of the App Store for the first time because when I check out these apps, they rarely have any App Store reviews.
The upshot of all these statistics and trends is App Store chaos of a magnitude that we haven’t seen in a long time. It’s a little like the App Store gold rush of the early days, but without the gold. Discoverability has gone from bad to worse, and the supply of apps is off the charts compared to the demand. With download numbers barely creeping up, the flood of new apps is making selling on the App Store harder for everyone.
Yet, it’s exactly this sort of chaotic disruption that leads to exciting new apps. And, it’s tools like Codex and Claude Code that empower and democratize app development, allowing people who would never have built an app to see their ideas become a reality. Those are things I love to see.
There’s no doubt that the App Store is out of whack thanks to agent-assisted coding, but like any market it will find its equilibrium again. In the meantime, it’s never been a better time to be a fan of apps because while a lot of those record numbers are comprised of mediocre apps, there are hidden gems, and we’re on the hunt for them.
Saron Yitbarek, writing on the WebKit blog:
In Safari Technology Preview 247, we’re introducing the Safari MCP server — a Model Context Protocol server for web developers that makes your web development and debugging workflow faster and more powerful. We know agents are increasingly integral to the coding process and the Safari MCP server gives your agent the ability to know how your code actually renders in the browser by connecting it to a Safari browser window.
Any MCP-compatible client can connect to the Safari MCP server. By connecting your agent to a Safari browser window, your agent can emulate what your users experience, giving it the information it needs to debug more autonomously, like access to the DOM, network requests, screenshots, and console output.
Importantly:
The Safari MCP server runs entirely on your local machine and makes no network calls of its own. It also does not have access to your personal information in Safari (e.g. AutoFill or other browser activity). When it captures page content, screenshots, or console logs, that data goes directly to the agent you’re running — not to Apple. What happens to that data from there depends on the agent and model you’re using. As with any agent you give access to your browser, only use ones you trust.
For the past few months, I’ve been using Google Chrome on my MacBook Pro and Mac Studio not because I like the browser (in fact, I really dislike Chrome’s text rendering and UI), but simply because it was the best option for agents. In Codex specifically, between Playwright, Chrome Dev Tools, and OpenAI’s own Chrome extension, I could kick off research tasks (such as vacation planning and booking a hotel) that involved a browser directly from my iPhone, letting Codex drive the research on my remote Mac.
Safari in Codex.
Now, thanks to Safari’s new MCP server, I no longer have to use Chrome on desktop and can return to a unified browser setup across all my devices. Even better: it actually looks like Apple shipped the most ergonomic browser MCP for agents to date. The MCP server has dedicated tools for page extraction (including getting webpages as Markdown, based on WebKit’s own conversion pipeline), evaluating JavaScript, DOM interactions (clicking, scrolling, resizing the viewport for mobile screens, etc.), taking screenshots, and more. I set this up immediately in Codex, and I also asked Codex about comparing Safari’s MCP to its own Chrome extension and the older Playwright. The verdict: although Chrome has a richer API with Chrome Dev Tools when it comes to network requests, Codex actually preferred Apple’s leaner, more direct approach for letting an agent drive and debug a browser.
I’m really happy to see folks at Apple embrace agentic tools: between the new MCP capabilities of Xcode and now this, it looks like Apple’s software (on the Mac, of course) is becoming more and more approachable by people who are working in new ways thanks to agents. Whether you’re a web developer or tinkerer, I highly recommend checking out what Apple has released in Safari Technology Preview. More of this, please.
My Codex setup.
In the current era of coding agents becoming productivity assistants, iPadOS’ limitations are no longer defined by the lack of desktop-class multitasking or access to external peripherals. A new class of iPadOS shortcomings looms large on the horizon: the iPad’s app sandboxing and the absence of an open filesystem have relegated it to acting as a remote control for agents.
I’ve been ruminating on this thought for some time. I, of course, am very satisfied with the improvements Apple brought to the iPadOS platform throughout the 26 cycle: between new windowing, local capture, a menu bar, and the improved Files app, I can get most of my old-school work done on the iPad Pro without begrudging the experience like I used to a few years ago. What I’ve noticed this year, however, is that iPadOS 26 was an update designed to address the pre-existing low-hanging fruit of quasi-desktop efficiency; by design – and we have Craig Federighi on the record saying as much – it wasn’t intended to reinvent the wheel in any meaningful way.
That would have been fine a few years ago. The problem is that the productivity wheel is literally being reinvented by other companies as we speak. And in an unsurprising turn of events, the Mac is at the forefront of that reinvention, while the iPad is being left behind once again.
I’m writing this as someone who likes to be on the bleeding edge of AI tools. Since the release of Claude Code in early 2025, I (and, realistically, millions of other nerds whose taste aligns with mine) have increasingly moved a lot of my research, automation, coding, and other tasks considered “productivity” to desktop agents. These are powered by the same underlying technology as chatbots, but they’re considerably more powerful: Claude Code and Cowork, Codex, Perplexity’s Personal Computer, and self-hosted agents like OpenClaw and Hermes can understand intent and perform long-running tasks on your computer. They can write code, stay on a task for hours (or even days) at a time, analyze the contents of your screen, and even use a computer on their own. To believe that these AI tools are still a fad and chatbots are useless is misinformation at this point.
As of June 2026, I spend most of my days with two apps constantly open on my M4 Pro MacBook Pro: Codex and Notion. I’m a huge fan of what OpenAI has accomplished with Codex for Mac; their all-in-one package that combines thread-based chats with skills, plugins, connectors, computer use, memories, and system-wide filesystem access has genuinely allowed me to produce content for my audience that I wouldn’t have even dreamed of years ago. Since I started working with Codex, I’ve released a major update to Apple Frames, two CLIs (with two more coming soon), Shortcuts Playground, and a brand new MacStories Shortcuts Archive. These are actual products that are used by thousands of people and are generating real revenue for MacStories. I’ve also created an entire suite of MacStories-only internal apps (now used by our whole team) and am currently in the process of researching my iOS 27 review, creating an app for Shortcuts Playground, and brainstorming a different product altogether for MCP.
I’m doing all of this on a Mac because the true Codex experience isn’t even “desktop-only”: it’s Mac-only.
Let me give you some examples. To make the original version of Shortcuts Playground, I had Codex with Computer Use scrape my entire library on the old MacStories Shortcuts Archive, download over 300 shortcuts, and learn my coding style. Once I had a plugin, I also had Codex create over 1,000 shortcuts on its own, install them in the Shortcuts app for Mac, and debug them visually to find issues. For the new Shortcuts Playground plugin and app, I’m going even further: Codex created over 3,500 shortcuts spanning the entire library of Shortcuts actions while simultaneously coordinating with other Codex threads – some of which are running on my Mac Studio – to build an app, train a local model for Core AI, and constantly iterate on the new harness behind the app. Meanwhile, I’m also spending time in Codex organizing sources for my iOS 27 review, managing my tasks with Reminders and my own RemCTL, and performing general research on any topic I think of. The net result of all this is that I’m shipping more things, making more money for MacStories, saving money thanks to our new internal tools, and – most importantly – learning a lot and having fun.
Working with Codex for Mac.
Some of the CLIs I’ve built. Two of them are coming soon.
In all of this, the iPad’s role is to be, once again, an accessory to the Mac – and a rather expensive one at that. I wrote last year about iPadOS’ third-party app problem, and I still think all of it applies today. What’s different in 2026 is that these powerful agentic tools that are redefining coding and knowledge work, and which are centralizing productivity inside so-called “super apps”, aren’t coming to the iPad, but not for political or financial reasons: they’re Mac-only because only macOS is mature and open enough to support them.
In a funny twist of fate, Apple spent years trying to make iPadOS catch up to macOS on the surface, only for macOS’ underlying nature to unlock a new generation of apps that are rewriting the rules of modern software. No amount of windowing and menu bars can change the fact that neither Claude Cowork nor Codex will ever be possible on the current version of iPadOS. From an open filesystem that lets agents poke around anywhere to powerful Accessibility and Screen Recording APIs for developers to features like system-wide hotkeys, dictation, and true background processes, Apple’s oldest operating system is the one that feels the freshest (and the most powerful) in the AI industry today.
In fact, the Mac is so powerful and well-equipped for this new wave of desktop agents, it even yields an amazing experience when used in a “headless” setup with no display attached to it. This may have started as a trend because of OpenClaw, but six months later, I believe that a Mac mini with Codex running on it is the best bang for your buck if you’re building an always-on agentic setup based on a frontier model that can do just about anything you throw at it.
Which brings me to the iPad’s only silver lining in this space right now.
Most of my Codex tasks are running on my Mac Studio server, but I also use Codex on my MacBook Pro. Thanks to Codex’s support for remote connections, I can chat with threads running on the Mac Studio from the Codex app on a different machine, which is a fantastic feature. Alternatively, I can use Apple’s excellent (and, again, Mac-only) Screen Sharing app to log into my headless Mac and interact with macOS with superior performance and image quality to VNC. (I also use this HDMI 2.1 dummy plug to unlock 4K resolution when headless.) Coordinating Codex threads across different projects and machines feels like the sort of thing that shouldn’t be possible, yet here we are.
If the iPad is still relevant to me in this specific new way of working, it’s entirely because of the ChatGPT app and the fact that OpenAI cared enough to design a good remote experience for iPhone and iPad users. The Codex section inside ChatGPT is a solid product and getting better every week; it’s also the only app of this kind that lets you connect to multiple computers at once and manage threads across all of them. With Codex remote on iPad, I can continue my projects from the MacBook Pro or Mac Studio, pin specific threads, and keep an eye on all my agents working in the background. In general, it’s also a way to bridge functionalities that are missing from iPadOS. Without Codex remote, I wouldn’t be able to ask my agent to reschedule my reminders and turn some emails into tasks in the same workflow; I wouldn’t be able to tell my agent to go to the Apple Newsroom, find me a specific press release, extract all images from it, and upload them to our CDN.
Codex remote on iPad.
Perhaps Siri AI will unlock similar capabilities (although I doubt to the degree of a frontier desktop agent), but it’ll still be limited to apps made by companies that decide to support App Intents. When I’m not using my Mac and I think of something I can’t do on the iPad these days, I instantly reach for Codex inside ChatGPT. I suspect this trend will only grow as more AI companies release similar remote experiences for iOS and iPadOS.
Here’s the issue, though: these are remote companion experiences. It’s great that they exist and that OpenAI cared enough to make them great, but they’re still no substitutes for the real Codex experience on a Mac. It’s ironic if you think about it, isn’t it? A headless Mac is still more flexible and powerful than an iPad hamstrung by the closed nature of its own OS and software gatekeeping.
My headless Mac Studio in Mirage for iPad.
Mirage for iPad’s app launcher.
If you’re among the five million people using Codex on a weekly basis and will be planning your next laptop purchase in a few months, perhaps these considerations will change the calculus for you. Especially in the face of price increases and RAM uncertainty, would you be better served by spending $1,499 on a hybrid laptop that gets work done the old way with 256 GB of storage, or $1,499 on a computer where AI is a first-class citizen with 24 GB of RAM and double the storage? If we accept that power users’ priorities and workflows will be dramatically changing in the near future (if they haven’t already!), all prices being equal, which OS is best positioned for the future?
Of course, I like using both an iPad Pro and a Mac for different tasks. But let’s face it; in this day and age – and especially after last week – not everybody has the budget for two computers that have some overlap between them. Sure, I could argue that an iPad Pro + a headless Mac mini is an insane combo to get the best of both worlds these days: the nimble and delightful nature of iPadOS and the freedom to experiment with AI on a Mac. You can install apps like the latest Jump Desktop beta or the great new screen sharing app Mirage on the iPad and use your Mac mini almost as if it were connected to the iPad’s display. Set up Tailscale on your network, and the Mac’s power can be with you at all times. But that’s a lot of money, and a lot of setup, to achieve an approximation of what a single MacBook Air can handle with more memory, more storage, and a couple of Mac apps.
I’m sure that Apple is aware of the changing nature of this landscape, but in a similar vein to chatbots and Siri, perhaps they were unprepared for the tide of agents to accelerate so quickly. If the idea of putting macOS on an iPad Pro continues to be a non-starter at Apple, and assuming agent-based coding and knowledge work are here to stay, what should the company do to make sure the iPad doesn’t get left behind again? Is a sandboxed Terminal for iPad backed by a dedicated “Agent Framework” the solution? Is working with other companies to build more “remote companion” apps the way forward?
Or, perhaps more obviously: how long until Siri AI can also become a capable agent, connected to App Intents, on par with competing frontier agents?
Until, and if, that happens, the iPad will once again play the role of sidekick to macOS.
When it comes to the bleeding edge of AI and pushing the frontier of agents forward, there’s nothing like the Mac.
The future of military artificial intelligence may not be a single all-knowing machine directing a battlefield from a distant command center. Instead, DARPA is preparing to explore enormous collectives of AI agents capable of organizing themselves, dividing up missions, adapting to chaos, and continuing to operate even when some members of the group fail, disappear, or go rogue.
The Defense Advanced Research Projects Agency (DARPA) has formally launched a new effort called Decentralized Artificial Intelligence by Controlled Emergence, or DICE, aimed at developing AI collectives that can operate in contested environments without requiring a human or central AI planner to explicitly direct every action.
By the program’s final phase, DARPA wants researchers to test simulated collectives involving as many as 100,000 AI agents exchanging up to 1 million messages as they work toward shared objectives.
A recently released solicitation notice describes DICE as a 36-month effort divided between decentralization, adversarial robustness, and large-scale testing.
“Future conflicts will unfold at machine speed, and the current centralized planning and scripted orchestration approach is too slow, too rigid, and too predictable for the hyper-dynamic, contested environments of the future,” researchers write. “Surprise, adaptability, and resilience will depend on the ability to compose and recompose AI capabilities on demand.”
The basic premise behind DARPA’s DICE program is that today’s approach to multi-agent AI has a serious weakness. Most systems involving multiple AI agents still rely on some form of central orchestrator.
Think of it as a manager overseeing a team of specialized employees. One AI agent might gather information. Another analyzes it. A third creates a plan. A fourth executes a particular task. But somewhere in the system, a central planner or workflow typically decides who does what and combines the results.
That works reasonably well for predictable tasks. However, DARPA argues it may be dangerously vulnerable on a battlefield.
A central orchestrator creates a bottleneck because every increasingly complex mission must ultimately flow through a single decision-making structure. As the number of agents, roles, and interactions grows, the central planner must absorb more information, preserve greater context, and coordinate an expanding web of specialized systems.
Eventually, DARPA says, even an AI orchestrator can hit the context and inference limits of its underlying foundation model. There is also a more obvious military problem. A central command node is a single point of failure.
If communications are degraded, information arrives late, or the orchestrator itself is compromised, errors can potentially spread throughout the entire AI collective.
In a future conflict where electronic warfare, cyberattacks, deception, and broken communications are expected rather than exceptional, that vulnerability could be catastrophic.
DICE proposes a radically different model.
Rather than having a central AI issue instructions, individual AI agents would communicate directly with one another. They could form temporary teams, divide missions into smaller tasks, negotiate roles, combine incomplete information, and reorganize themselves when conditions change.
DARPA compares the idea to the architecture of the internet, where resilient global behavior can emerge from relatively simple local rules and decentralized interactions.
Essentially, an AI agent might receive a mission and determine that only part of the objective falls within its capabilities. The agent could volunteer to handle that portion and broadcast the remaining problem to neighboring agents.
Another agent might take the next task. Others could compete to perform the same job, potentially bidding based on cost or capability. Through repeated local interactions along with distributed consensus, the collective could gradually assemble its own mission plan.
If an agent fails, the entire mission does not necessarily need to be replanned. Its particular task could simply be offered back to the collective.
DARPA calls the larger concept “controlled emergence.”
Similar systems already exist in nature. Ant colonies, or flocks of birds, can engage in surprisingly complex group behavior even though no individual member possesses a complete understanding of the collective.
Computer scientists have long studied similar phenomena. DARPA’s solicitation even points to Conway’s Game of Life, the famous mathematical simulation in which extraordinarily complex patterns can emerge from a handful of simple rules.
The challenge is applying that principle to cognitive systems capable of reasoning, planning, and independently pursuing goals.
A sophisticated AI system may interpret ambiguous information, interact with tools, maintain memory, and generate its own intermediate objectives. Put thousands of such agents together, allow them to communicate, and the resulting collective could behave in ways that were never explicitly programmed.
That unpredictability is both the attraction and the danger of DICE.
DARPA wants the collective to remain creative enough to discover unexpected courses of action. However, it also wants individual agents to continue aligned with their designated roles, the wider mission, established doctrine, and ultimately the human user’s intent.
Too much control could produce a predictable but unstable AI force incapable of adapting. Too little control could allow agents to drift from their roles or develop what DARPA calls “instrumental goals” that are not consistent with the mission.
In other words, the Pentagon’s research arm wants emergence—but only within boundaries.
To solve the problem, DICE researchers will develop a local adaptor designed to interface with each AI agent. Rather than functioning as a central command system, each adaptor would operate locally, helping the agent coordinate with peers while simultaneously monitoring and modifying its behavior.
For AI models whose internal activations are accessible, researchers may experiment with techniques such as “activation steering.”
The simplified idea is that assigning an AI a particular role can create detectable patterns inside the model. Researchers may be able to identify a mathematical representation associated with that role and continuously steer the model back toward it when the agent begins to drift. DARPA describes these representations as potential “role vectors.”
For proprietary or black-box AI models whose internal activations cannot be directly examined, the challenge becomes considerably harder. DICE researchers may instead experiment with memory editing, context engineering, tool restrictions, or even game-theoretic incentives created to encourage collaborative behavior.
The solicitation draws an intriguing comparison to the way social norms emerge and spread among humans. As AI models become better at reasoning and more aware of the context in which they operate, DARPA suggests that the same growing sophistication that fuels fears about rogue AI could also provide a new way to control it.
The scale DARPA ultimately envisions is enormous.
During the first nine-month phase of DICE, simulated missions are expected to involve approximately 500 agents and 5,000 interactions. DARPA will compare the decentralized systems with state-of-the-art centrally orchestrated multi-agent AI systems.
The second phase increases the target to 5,000 agents and 50,000 interactions while introducing adversarial conditions. Agents may fail, receive deceptive information, or deliberately behave as compromised or rogue members of the collective.
By the final 12-month phase, DICE aims to test simulations involving 100,000 agents and 1 million interactions.
At that scale, DARPA is not simply studying better chatbots working together. The agency is exploring whether collective intelligence itself can become a military capability.
One example described in the solicitation imagines AI agents conducting operational planning across an entire conflict ecosystem. Some agents could negotiate for raw materials in contested areas. Others might manage supply chains, optimize manufacturing, deploy units, conduct battle management, or analyze post-battle damage.
Even an individual simulated drone could effectively become a miniature AI collective composed of separate navigation, perception, and networking agents.
Under the DICE concept, those capabilities may potentially be recombined as conditions change. DARPA offers the example of a surveillance drone being repurposed as a network router because the collective recognizes an emerging communications need.
The important distinction is that no central planner necessarily has to foresee that solution. The collective discovers it.
Still, DARPA is careful to note that DICE is a research program centered on simulation. The agency’s public program page explicitly states that its scope does not include developing or launching autonomous systems in the real world.
Instead, DICE will use Department of War-relevant simulated environments to measure whether decentralized AI collectives can outperform today’s centrally orchestrated systems. Those simulations may resemble extraordinarily large and complicated strategy games.
DARPA suggests researchers could draw inspiration from StarCraft or massively multiplayer online role-playing games, although the environments must remain relevant to military problems. High-fidelity physics is not necessarily required. In fact, rendering every drone, vehicle, or battlefield interaction in photorealistic detail could make a simulation of 100,000 AI agents computationally impossible.
The focus is on reasoning. DARPA wants to know whether thousands of AI agents can gather and combine fragmented information without simply reinforcing one another’s incorrect assumptions. The systems will be tested on their ability to recognize deception, isolate compromised agents, adapt when communications fail, and remain focused on a mission across thousands of reasoning steps.
Perhaps most importantly, DICE is designed to examine how these massive AI collectives behave when their interactions yield outcomes their designers never anticipated.
DARPA explicitly wants its test environments to provoke emergent collaboration, competition, and even collusion.
During later phases, DICE teams will face one another in simulated “team vs team playoffs.” The second phase will culminate in a competition used to select which research teams advance. The final phase will feature another head-to-head competition after the systems have been pushed toward full scale.
By Phase 3, the program will bring in another significant complication: AI agents creating other AI agents. That progression draws attention to the deeper tension at the center of DICE.
For years, much of the AI safety debate has focused on preventing autonomous systems from exhibiting unforeseen behavior. DARPA is approaching the problem from a different direction. The agency appears to believe that, in future contested environments, some degree of unforeseen behavior may be precisely what gives an AI collective its value.
A perfectly scripted system is predictable not only to its commanders but potentially to an adversary.
DICE is therefore attempting something considerably more ambitious than building a larger multi-agent AI framework. DARPA wants to understand whether engineers can deliberately cultivate emergent machine intelligence, allow it enough freedom to improvise, and still build controls capable of keeping an enormous AI collective aligned with a human-defined mission.
Whether those competing demands can coexist remains an open question. The program’s own solicitation repeatedly acknowledges the possibility of role drift, deceptive agents, collusion, compromised systems, and AI agents independently developing misaligned actions.
Ultimately, those are not side issues. They are the fundamental questions the DICE program was created to answer.
“DICE is built for such autonomy-powered, high-tempo battlespace,” the solicitation notice reads. “It is a future where success is measured by the relentless, creative ability of an autonomous collective to execute missions in the face of chaos and uncertainty. This is a profound and necessary departure from the past.”
Tim McMillan is a retired law enforcement executive, investigative reporter and co-founder of The Debrief. His writing typically focuses on defense, national security, the Intelligence Community and topics related to psychology. You can follow Tim on Twitter: @LtTimMcMillan. Tim can be reached by email: tim@thedebrief.org or through encrypted email: LtTimMcMillan@protonmail.com
What happens when you give AI coding agents a lab full of robotic arms, some compute resources, and a “generous token budget” for teaching the robots various tasks? The agents can apparently figure out a training regimen that teaches the robots to successfully cut zip ties and even insert GPUs into thin sockets on motherboards.
That glimpse into how AI can act in a fully autonomous way to automate robot training was made possible by a new agent harness framework—software that wraps around AI models to enable their use of various tools while also providing capabilities such as memory, context, constraint, and feedback loops. That agentic harness, called ENPIRE, was developed by robotics researchers at the Nvidia GEAR (Generalist Embodied Agent Research) lab alongside collaborators from Carnegie Mellon University in Pittsburgh and the University of California, Berkeley.
“A part of our NVIDIA GEAR lab now self-improves tirelessly overnight,” wrote Jim Fan, director of AI at NVIDIA, in a LinkedIn post. “We just read the reports in the morning.”
© NVIDIA
Login