Six takeaways from OpenAI's DevDay: agents emerge as the backbone of digital life

OpenAI's DevDay produced more than twenty announcements, but the underlying message is simple: the company wants the user's entire digital life to run through ChatGPT and through agents capable of carrying out tasks on their behalf. With 1.2 billion weekly users as its distribution base, OpenAI is entering a personal-agent race that already had competitors under way, such as Muse and Instinct (with more than 500,000 and 100,000 trial users respectively, according to The Information) and GrokBot. Its answer is called Dots, a personal agent based on the GPT-6 Astra model, with its own "cloud computer" and a plugin ecosystem that connects to more than 4,000 applications. The promise is an assistant that runs continuously, learns preferences and takes on recurring responsibilities; the article cites as an example a test case in which an agent detected a pending invoice from a collaborator, prepared it and sent it after receiving approval.

The underlying idea is for the agent to become the "organizing principle" of digital life: instead of the user remembering where everything is and carrying information between applications, the agent does it. Whoever controls that agent gains influence over which services are activated next, much as Google built its business around the search box; now personal-agent companies aim to sit alongside the person, all day long, free of charge for the time being.

A second set of announcements illustrates the speed with which OpenAI replicates moves by smaller startups. TypeSafe AI had launched its product Jev on 15 September, and barely two weeks later OpenAI unveiled its own Decisions API, based on a model called Luna, to answer questions, interpret images with predefined options, classify information and route an agent's next actions, for now in limited preview. Jev had already shown real traction: according to Vercel data cited in the article, close to 13% of the paying teams on its AI Gateway adopted it within the first 24 hours, the fastest adoption rate recorded on that platform. The author does not claim that OpenAI has surpassed Jev, nor does he know when it began developing its alternative, but he stresses what he considers a lesson of the current moment: identifying a new category no longer guarantees room to manoeuvre, because a successful launch can, in practice, turn into a very well-documented feature request for the entire industry.

The third thread, according to the article, debunks the narrative that AI was going to wipe software companies off the map (the supposed "SaaSpocalypse"). OpenAI Marketplace allows certain enterprise customers to apply part of their existing spending commitment with OpenAI to approved partner products — Salesforce, ServiceNow, Adobe and Figma are mentioned — although the customer still contracts and pays those partners directly, while OpenAI reconciles the eligible spending against the commitment. Added to this are plugin extensions that give those applications a place inside ChatGPT: sidebars, panels next to conversations and interfaces for viewing and editing files. The reasoning is that replicating a visual interface is only part of replacing payroll systems, customer records or legal workflows built up over years; agents can make them easier to use and increase the volume of work that flows through them, but they do not necessarily replace them. The nuance, the author notes, is that if the customer increasingly reaches that software through a third party's agent, that third party gains influence over which tool is chosen and how it is compared with alternatives, even if the underlying software survives.

A fourth announcement, probably the most noticeable for users in the short term, is speed. The new "Ultrafast" tier promises up to eight times faster token generation in Codex (up to 300 tokens per second) and up to six times faster in the API, although this measures generation speed, not that each query is resolved eight times faster overall. The central argument is psychological: once users get used to near-instant responses, any slower interaction starts to feel broken, just as happened historically with the jump from dial-up to broadband or from slow phones to responsive smartphones. That "zero wait" expectation may prove especially uncomfortable for companies left on the slow side of that transition, particularly when combined with the low switching costs that a highly adaptable AI makes possible.

The fifth point, described as the least flashy feature but the one with the greatest practical impact, is the ability to connect multiple email, calendar and contact accounts (already available for several Gmail, Google Calendar and Google Contacts accounts before DevDay, but more relevant now that the agent aims to understand the user's entire life). Most people spread their lives across several inboxes — work, personal, an old account that still receives something important — and today have to manually reconstruct the context scattered among them. An agent with access to the relevant accounts can anticipate conflicts, for example between a work invitation and a personal flight booking, before the user mistakenly commits. The article does warn, however, that being able to consult personal and professional information should not automatically mean being able to share one with the other, which makes data loss prevention (DLP) a serious concern for corporate security teams.

Finally, the author argues that the mobile playing field is in play again for the first time in decades, without OpenAI having presented a phone of its own. Dots can be used from ChatGPT on mobile, desktop and web, as well as from Slack and Teams, with text-message support planned, keeping context across channels and carrying out the work on its own cloud computer. If users increasingly just explain what they want and approve the result, much of the intermediate work happens in software they never open directly, and the agent gains influence over which app or service receives each assignment. The article qualifies that Apple and Google still control very valuable pieces of that experience — operating system, distribution, device permissions — and that a good demo does not amount to a surrender by Cupertino, but it warns that OpenAI does not need to replace iOS or Android to change the economics of mobile: it is enough to become the usual starting point for the user's attention.

Taken together, the article concludes that these pieces fit with one another: the personal agent knows the user, a decision model helps it choose what to do, existing software provides the specific capabilities and greater speed makes the whole interaction feel natural, while connected inboxes supply context and mobile offers the access channel. The result, according to the author, is considerable ambition for a company that many people still associate only with a chat box: going from showing us where to click to becoming the entity we simply ask, so that we do not have to do it ourselves.