To a packed house of developers, Sam Altman and crew released a passel of new products and capabilities, thrilling the faithful and serving notice that OpenAI is moving quickly toward its own particular Altmanesque vision of the agentic future. There were more than 20 announcements, which is a considerable number of things to digest in one sitting, but the underlying proposition was reasonably simple and utterly predictable. That is, bring your digital life into ChatGPT and let our smart agents figure out how to GSD. OpenAI says it now reaches 1.2 billion weekly users, giving it a substantial audience for this proposition and, as we know, distribution is even more a sport of kings in the AI Era. Most of those users for now are ChatGPT to date but it’s a certifiable juggernaut. That said, the real question is how much of the surrounding technology industry gets invited along, how much of the way we work and use tech gets rearranged and who gets to collect the money once our agents take the wheel for real. Here are six quick and dirty takeaways.
Agents Are Officially Now the Center of Gravity in Consumer Tech
It’s official. True, OpenAI is arriving at a party where other guests have already started rearranging the furniture. Muse and Instinct have taken off: The Information reported more than 500,000 Muse trial users and over 100,000 Instinct users despite limited availability. GrokBot has also been winning enthusiastic reviews from some early users. Enter Dots, OpenAI’s entrant in the personal agent space. Powered by GPT-6 Astra, equipped with their own cloud computers and connected through a plugin ecosystem spanning more than 4,000 applications, each Dot becomes the primary touchpoint for a user accessing the increasingly sprawling ChatGPT firmament. OpenAI describes agents that work around the clock, learn your preferences and take on ongoing responsibilities. TBD but, early user testing (mine) gives them a thumbs up on this one. Proactive insertions are winning kudos. One early tester’s Dot noticed that he had forgotten to invoice a publication, prepared the invoice and sent it after approval. As somebody who writes for publications and sometimes speaks for cash, I consider remembering to collect the money a respectable contribution to the advancement of intelligence.
The larger implication? We can now see how the agent could become the organizing principle of your digital life. We have spent decades learning where software companies put things, remembering passwords and carrying little bundles of information between applications like digital nomads hauling our belongings on our backs. Now we are being offered something that can do the carrying and not only carry the water but remind us when we are about to run out and where the next well is and when to buy our mom flowers for her birthday. (deep breath). Like Muse and Instinct, Dots is making a play for our preferences. Once it knows your predilections, your obligations and which airline you would rather walk than fly, you have a reason to keep using it, of course, although agent promiscuity is likely to be a thing. Whoever operates that agent acquires influence over which services get called upon next. Google built an extraordinary business standing beside the search box. The personal-agent companies would like to stand beside the person, all day, and be your concierge and your butler. Thus far, for free.
OpenAI Decisions and Jev: First Mover = Even Faster Follower
TypeSafe AI introduced Jev on September 15. Two weeks later, OpenAI announced its Decisions API, using its Luna model to answer questions and parse images with predefined choices, classify information, route requests and select an agent’s next action. It launched in limited preview, tapping directly into the contrails of Jev. Product market fit is best when someone else does it for you, right? Jev had already demonstrated substantial demand: Vercel reported that nearly 13% of paid AI Gateway teams used it within its first 24 hours on the service, making it the gateway’s fastest-adopted model. That OpenAI would fast follow is not a surprise. The real surprise is how fast follow the fast follow was, and what that means for any startup edge that might be accrued, even through a legendary launch like Jev’s.
To be clear, we don’t know when OpenAI began developing its offering, and a product announcement does not establish that it has beaten Jev. But the competitive predicament is familiar and the speed is striking. A startup identifies something useful that the giants have overlooked, developers rush in, and suddenly one of the giants has a product in the same neighborhood. The startup gets credit for seeing the opportunity. The giant gets to offer its version through accounts, billing arrangements and developer relationships that already exist.
This is approximately the uncomfortable moment in the entrepreneurial journey when the founder discovers that validation of the market and an extremely large competitor can arrive in nearly the same hype cycle. (The rapid upwelling of dozens of open source Jev clones certainly smoothed the path). Jev may still win on quality, speed, price or simply being the tool developers prefer. What this episode makes difficult to believe is that identifying a new category buys much breathing room in the agentic age. In AI, even a successful launch can amount to sending the entire industry a particularly well-documented feature request.
The SaaSPocalypse Is Officially Over
Apparently, the software companies are coming to their own funeral to fund the pupus and drinks. OpenAI Marketplace lets eligible enterprise customers apply part of their existing OpenAI commitment toward approved partner products, with Salesforce, ServiceNow, Adobe and Figma among the companies on the roster. Customers still contract with and pay the partners directly, while OpenAI reconciles eligible spending against the commitment. Meanwhile, plugin extensions give software companies places inside ChatGPT to put their applications, including sidebar entries, panels alongside conversations and interfaces for viewing and editing files. The industry that AI was supposedly going to vaporize has been offered distribution, integration and a way to participate in the budget. The system of record gets the big W, and trusted data stores remain locked in place by the same forces that have effectively frozen IT departments in blocks of ice for millenia. (I exaggerate, to be fair).
Taking off the rose-colored AR glasses, this only makes sense when you consider what businesses actually buy from SaaS. A company uses payroll software because employees have an unreasonable attachment to receiving the correct amount of money on the correct day — a pretty non-deterministic task. Its customer records, financial systems and legal workflows come with years of accumulated rules, integrations and institutional knowledge. Generating a plausible interface is only one part of replacing any of that. Agents could make those systems easier to use and increase the amount of work passing through them, even as they put pressure on particular products and pricing models. (and, to be clear, both Anthropic and OpenAI have gone a long way down that path with Gmail, Slack, Google Calendar and other integrations that allow us to invoke these business tools with our Dots or chats). However, ceclaring the entire category dead always required skipping over a substantial amount of inconvenient detail.
There is, however, a catch (ah, there’s always a catch). If the customer increasingly reaches your software through somebody else’s agent, that somebody else may gain considerable influence over your business. The agent can choose tools, compare alternatives and keep the user happily occupied without ever sending them to your homepage. Software survives, but the relationship changes. The hotel still has the rooms after a booking platform arrives. It may simply find that somebody else has acquired a profitable interest in deciding who sleeps in them and might play some role in routing them to a different hotel over the fullness of time.
Soon Consumers Will Learn To Expect “Zero-Wait” AI and Agents
The announcement I suspect people will feel most immediately is speed. OpenAI’s Ultrafast tier promises up to eight times faster token generation in Codex, reaching 300 tokens per second, and up to six times faster generation in the API. That is a measure of generation speed, rather than a promise that every query finishes eight times faster. Sure, an agent can still spend time waiting for a website or retrieving information.
But speed does something to our perception that a pricing table cannot quite capture. I’ve watched mesmerized as a Cerberas monster silicon system rendered browser game development and imaging in near real time by barfing out tokens at an ungodly pace. Such speed thrills but its not on offer broadly. This is why we have the AI latency issue. Ask an agent to tell you if the train you are about to get on in Tokyo is going in the right direction, and you don’t have 10 seconds to wait. Ask the same question and receive a useful answer almost immediately, and you get on the train with confidence and continue your conversation, expecting real-time all the way down. The interaction starts to feel more like working with something that is present. You make an adjustment, it responds, you notice something else and keep going. We are conversational creatures, and a conversation in which the other party disappears for ninety seconds after every observation is generally interpreted as a problem.
Once we get used to that responsiveness, everything slower is going to feel horrible. This has happened before. Broadband made dial-up unbearable. 3G made 2G feel terrible, rinse and repeat to 5G today. And responsive smartphones like the iPhone made sluggish, limited phone interfaces feel broken. Nobody congratulates a modern website for loading in less than a minute, even though that would once have been a perfectly respectable performance. AI could compress that adjustment into an especially unpleasant period for companies on the wrong side of it, particularly with the reduced switching costs of highly adaptive AI models. (You can not only fork software but you can fork your life more easily with AI, too!)
Faster agents will also make the software we operate ourselves feel increasingly cumbersome. After telling an agent what you want and getting it back quickly — in real-time conversational turns — being asked to navigate six menus and export a CSV begins to feel like the computer has assigned you homework.
“Bring Your Whole Email Self” Is the Killer Agent and AI Feature
My candidate for the least glamorous feature with an outsized effect on usefulness is the ability to connect multiple email accounts. ChatGPT (and others) already support multiple Gmail, Google Calendar and Google Contacts accounts, a capability that predates DevDay but becomes more important when an agent is supposed to understand your life. Most of us are distributed across a small collection of inboxes. There is the work account, the personal account, the other work account and an elderly address that should have been retired years ago but still receives something essential twice a year. Software regards these as separate identities. Unfortunately, all of them expect the same human being to show up.
Suppose I want to know whether I can attend a meeting. The invitation is in my work account, the flight reservation is in my personal account and a relevant conversation took place in a third inbox because somebody used the address their autocomplete happened to remember. I can assemble the answer myself, and have been doing so for years, painfully. My life is complicated, my history sharded, and my tolerance for this artificial severance of work and life has grown short. (See above). Enter the multimail verse! An agent with access to the relevant accounts has a chance to see the conflict before I agree to be in San Francisco while sitting on an airplane over Greenland.
To be clear, being allowed to consult personal and professional information should not automatically mean being allowed to share one with the other. DLP becomes a pretty serious concern for IT and security teams, but hey, that’s their job. For users, FINALLY connecting the relevant context to all of our selves is how personal agents become personal.
The Mobile Deck Is in Play for the First Time in Decades
The most interesting mobile implication of DevDay did not require OpenAI to unveil a phone. Dots can be reached through ChatGPT on mobile, desktop and the web, as well as Slack and Teams, with texting planned. Their context carries across those channels, and their work can happen on their own cloud computer. Follow that arrangement far enough and the device in your pocket starts to play a different role. You use it to explain what you want, look at the result and approve whatever requires your attention. A growing share of the intervening work happens somewhere else, in software you may never open. But let’s be real. Our phones rule our lives, With Muse, Instinct and others becoming “everything” apps for our mobile lives, their is a sea change afoot in who controls what we see, do and use on mobile. The deck is in play and that’s huge.
Apple and Google still control enormously valuable pieces of this experience, including operating systems, distribution and device permissions. They are also quite capable of recognizing a threat and responding to it. Nobody should confuse a good demonstration with the impending surrender of Cupertino. Apple Intelligence has also been showing signs of life lately. But OpenAI does not necessarily need to replace iOS or Android to change the economics of mobile computing. It needs to become the place where enough people begin. If I ask my agent to arrange dinner, find a flight or deal with an outstanding bill, the agent has influence over which application or service gets the assignment. The icons can remain exactly where I left them while the business of directing my attention moves elsewhere.
Conclusion: Your Agent Mileage May Vary But The Wheels Are Spinning
To tie it all up with a bow, the personal agent gets to know you, the decision model helps it choose what to do, existing software supplies capabilities, and faster intelligence makes the whole arrangement feel increasingly natural. Your assorted inboxes give it the context to be useful. Your phone gives you a convenient way to reach it. OpenAI would like to put itself in the middle of those relationships, which is a considerable ambition for a company many people still think of as operating a chat box. For thirty years, technology companies taught us where to click. Now they are competing to become the thing we ask so we don’t have to. The monkey has acquired an assistant, and everybody would like to employ it. Me included.
