The Frog, the Reins and the Shop Front: OpenAI’s Bid to Own the Working Day

Whoever owns the friendly face in your Slack channel owns the daily conversation, and on Tuesday OpenAI spent a whole keynote trying to make that face its own.

It began, at 10:13 in the morning and in front of about 2,500 people in San Francisco, with silence. The demonstration of dots, the always-on agents that headlined DevDay, was meant to show an assistant that works while you are elsewhere. What it showed, for a moment, was one that was elsewhere. Simon Willison, blogging live from the room, wrote “still checking”, and then, in a phrase I rather admire, “an embarrassing silent moment”.

I do not say this to be unkind. Live demonstrations fail; that is why they are called live. I say it because the pause is a useful place to stand while looking at what OpenAI actually launched, which is less a new capability than a new arrangement. There is a face (the dot), a set of reins (the machinery underneath, offered to developers), a wallet (the ChatGPT subscription, now spendable elsewhere) and a shop front (a place in ChatGPT where other people’s software can set up). Taken one at a time they are sensible. Taken together they describe a company that would like to be the place where work happens, not merely the supplier of the clever animal that does some of it.

A pet, not a program

Start with the face. A dot, according to OpenAI’s own page, is a cloud-based agent that works toward goals around the clock, contacts you when it needs input, and learns your preferences. It runs on GPT-6 Astra and has its own cloud computer, with a browser, a terminal and file storage. You can open it, watch it, and press a button marked “Take over”, The Decoder reports. It connects to “over 4,000 apps”, and you can reach it in the ChatGPT apps, in Slack and in Microsoft Teams. Texting is “coming soon”.

This is sensible. It is also, said TechCrunch, something much of which was already possible through Codex, OpenAI’s coding assistant, and similar tools that wrap a model in helpful scaffolding. Dots bundle it, TechCrunch argued, with a focus on independent action. Willison remarked that the interface “really does look like Meta’s Muse agent”.

So the bet is not technical. It is about how an agent looks when it sits down in your office. Each dot is given a customisable cartoon face: a frog, a rabbit, a cloud, a “pet”. The demo presenter, Holly Li, had one called Dottie, with access to Slack, Teams and ChatGPT. It is a cheerful bit of design, and I do not think it an accident. Software that acts in your name, reads your calendar and answers your colleagues sounds faintly alarming when described plainly. Give it a frog and it becomes a colleague you might bring biscuits for.

Nor is the idea OpenAI’s alone. Meta’s Muse launched about three weeks earlier and is technically similar, The Decoder says, and Axios reports that it “has gone viral and moved markets”. OpenAI has looked at the shape that worked and drawn its own version, with better manners and a paid entrance.

That last part matters. Muse is free. Dots go, for now, to Pro and Business Premium customers, and the analyst Shay Boloor of Futurum Equities argues, according to the eChai roundup, that this gives Meta “a massive head start”. Pro users in the EEA, Switzerland and the UK are excluded for now, per OpenAI’s help page. There is also a wrinkle over who exactly gets one. OpenAI’s Tibo says dots are included in every Pro plan, including the $100 one, and Engadget agrees, but BGR says they are not available at $100. On this point I side with OpenAI and Engadget, and a pre-keynote report of a $200 requirement may explain BGR’s slip. It remains a disputed point, and I leave it so.

What I find more persuasive than the avatar is a modest example from OpenAI’s own account. An early tester’s dot noticed that he had forgotten to invoice a publication. It prepared the invoice. It sent it only after he approved. This is the actual product, the software equivalent of a good junior: it spots the thing you missed, does the drudgery, and then asks. OpenAI’s Sarah Friar called a dot “the agent that can tap you on the shoulder” when she spoke to CNBC, and that is the phrase I would keep. The controls follow the same instinct: rules for when to act and when to ask, and read-only tools while the dot is idle.

Carefulness does not settle the matter. WIRED cautions that agents can be tricked by adversaries or share information unintentionally, and that if your account allows model training, the setting carries over to the agent. And Willison, when he tried the link to create his own dot at 11:32, was told to switch to a desktop computer. You cannot make one on a phone at all, according to the help page, which is a curious limit for a thing meant to slip into daily life.

The reins

A face needs something behind it. The clever animal is the language model; the rig is everything else that lets it do a job: a set of tools it may pick up, a memory it can trim when the conversation grows long, a means of recovering when something goes wrong halfway. It is unglamorous, and it is increasingly what people mean when they say “agent” rather than “chatbot”. The interesting thing about DevDay is how many places OpenAI announced that its rig could now sit.

Codex got a refreshed command-line version, the plain text window where programmers work, in which you can start and steer tasks by voice, per the company’s recap. It can also run in the cloud, controlled from a phone, with reusable environments so a whole team starts from the same approved setup. The Decoder notes that phone control has existed since May; the shared environments are new. Tibo said in a post quoted by eChai: “it’s impossible to go back to building on your laptop once you’ve taken the time to configure it.” That is marketing, and also a good description of why people stay with a tool.

The dot sits in the same house: each gets its own cloud computer and can, with permission, reach into your laptop to run Codex there. Then there is the room developers can move into. OpenAI’s Agents API runs a managed version of Codex’s machinery for you, handling session management, context trimming and recovery, and it now supports computer use, meaning the agent can drive a web browser hosted by OpenAI. Beyond the API, OpenAI said it had worked with Amazon on Bedrock Managed Agents, so that OpenAI agents can run entirely inside AWS. A terminal, a cloud sandbox, a personal agent, a hosted service and a rival’s data centre: one house, several doors.

The commercial logic is not difficult. A model can be swapped, like a horse. Developers will move to whichever one is cheaper, faster or simply better that month. A rig is stickier. Once your team’s environments are configured and your approvals wired in, leaving means starting again.

I should be honest that the tidy picture is partly my own drawing, and the sources do not agree on dates. The Decoder says the Agents API has been in public beta since 11 September, another writer says the 10th, and OpenAI’s changelog files it under the 15th. The Bedrock preview is dated April by The Decoder and to now by TNW. And OpenAI’s own recap says the Agents API is available “in Codex and ChatGPT Work on Pro 500 and Enterprise”, a sentence that looks to me like it has wandered in from another announcement. Willison, for his part, had a bad experience with Codex in the cloud two hours before the keynote. That is one man on one afternoon, but a rig is only as good as the worst hour you spend inside it.

The wallet, and the shop next door

“I’ve wanted this one for years!” wrote Willison, in the middle of a keynote that contained, by OpenAI’s own count, more than 20 major announcements. He was not talking about the frog. He was talking about a button.

The button is “Sign in with ChatGPT”, and what is new about it is what sits behind it. According to OpenAI’s recap, users of Plus and Pro can now spend their ChatGPT plan allowance inside partner tools, across 16 partners, and set a cap on each. The recap names six: Cognition’s Devin, Notion, Vercel, T3, OpenClaw and Dactyl. TNW adds that users can cap each app’s weekly usage, and that the apps do not get access to conversations or memories. Engadget’s summary is the plainest: there is “no need to first set up an API key”. That looks dull. It is the most interesting thing said all day.

To see why, think about how a person uses an AI tool made by someone other than OpenAI. Either the developer paid for the model use and passed the cost on, which is expensive for a small developer, or the user generated an API key (a long password that lets a program use the model and meter the bill) and pasted it in, a chore most people quite reasonably decline. A subscription is different. A great many people already have one, and it already contains an allowance. OpenAI has let that allowance travel, rather as if the electricity you pay for at home could be used at a friend’s house, with a cap you set on the plug.

The identity-only part is not new: a beta of it dates to 29 July, with Airtable, GitLab, HubSpot, Notion, Supabase and Vercel, according to Magica and the contextos blog. What DevDay added is plan usage, which is to say the money. The timing is not coy: on the same day OpenAI cut the allowance on its $200 Pro plan and created a $500 one, so allowances have become something people will look at closely. My own inference, not any outlet’s claim, is that developers may come to see subscriptions as a cheaper source of tokens. I found nobody critical on this, and Enterprise administrators, Magica reports, can disable sign-in.

Then comes the shop front. The centrepiece is what OpenAI calls plugin extensions. In the company’s words, “We’re opening the platform we use to build ChatGPT features.” A developer can give a plugin a permanent place in the sidebar of ChatGPT or Codex, build interactive panels beside the conversation, and even write viewers for particular kinds of file. Altman, speaking to CNBC, described them as “essentially entire applications that feel native to ChatGPT”, and added: “you can build an editor, a dashboard, a whole workspace directly into ChatGPT and Codex.” That is a long way from a plugin as a tool the chatbot might reach for now and then. This is a shop unit with its own door.

According to Superpower Daily’s coverage of the keynote, Altman said Adobe and Figma were among “dozens of launch partners”, and TNW lists Canva, Figma and Adobe as early examples. That is the sensible way to launch a marketplace: famous tenants first, smaller shops following the footfall. OpenAI also promised a plugin creator, a redesigned submission process and better ranking and recommendations, with users choosing which plugins to install and approving what each may reach. A community member’s forum summary adds review tracking and human review requests; that is a single summary. OpenAI says ChatGPT now supports a proposed standard called MCP Events, which lets a plugin start something when an event occurs in a connected app; note the word “proposed”, for it is a specification still being argued over. And Sites, OpenAI’s website-building product, can now host plugins for Business, Enterprise, Healthcare and Edu customers. Willison’s notes credit Sites with reportedly 8 million sites, a figure from one attendee’s notes that OpenAI has not confirmed.

The third piece is the back office. The new OpenAI Marketplace lets eligible enterprise customers apply part of what they have already committed to spend with OpenAI toward approved partner software, from 32 first partners including Adobe, Figma, Salesforce, ServiceNow and Harvey. TNW says eligibility is limited to US enterprises and calls the partners a beta cohort. Put the sidebar, the plan-spending and the marketplace side by side and the design shows through. Developers get a place to be found, individuals a way to pay with a subscription they already have, companies a way to buy software with money already promised. OpenAI sits at each junction.

Now for the difficulty. The audience is real: OpenAI’s recap puts ChatGPT at 1.2 billion weekly users, which is the company’s own number, but a very large pool of people who already have an account and a habit. Yet the interesting question, whether developers actually want to move in, has almost no evidence behind it yet. I found no hostile coverage, and no developer reaction in the Hacker News threads I could see.

Willison’s live blog records Altman saying, in effect, that opening ChatGPT to outside developers had been tried before, with variable results. That is Willison’s paraphrase, not a verbatim quotation, so I have kept it out of quotation marks. I read it as a nod to earlier efforts such as the GPT Store and the Apps SDK, but that is only a reading. It is nonetheless the most honest line of the day. Why might this attempt differ? Not because the pitch is better; “build inside ChatGPT” has been made before. It is because this one solves the dull problem first. A store fails when the shopper has to do something awkward to buy, and a key, a sign-up and a separate bill are the awkward things. Sign in, spend the allowance you already have, set a cap: the reader can see themselves doing that on a wet Tuesday.

There is a tension the coverage has not touched, and it is my own observation. Notion is one of the partners you can now pay for with your ChatGPT plan, and also, as one Hacker News commenter put it about the new shared Space and Pages, a thing people wondered whether they were being told to replace: “are they telling us to use it instead of Notion or what…”. A landlord who runs a rival shop in the same arcade is not unknown, but the tenants tend to notice. And only six of the 16 partners are named on OpenAI’s page, the count being the company’s own. It is an announcement, not a result.

So return to the frog, and to the silence at 10:13. A silence in a demo is forgiven by lunchtime. A silence from an agent you have trusted with your invoices is a different matter, and it is one the frog will have to explain. The face may win the conversation, the reins may win the developers, and the button may win the wallet. But the models will keep getting cheaper and, one week at a time, more alike, and reins are what a hand gets used to. If I were asked which of Tuesday’s announcements would look largest in a year, I would pick the little button that lets one subscription pay for the other rooms in the house. Whether the 1.2 billion people on the pavement ever come in is a question OpenAI can answer only by keeping the door open, and the lights on, for longer than it managed before.

Sources: