OpenAI and Meta are packaging AI agents so more people can hand them work

1 · Decide who may switch on an agent with its own login. Owner: your chief information officer and the business owner of the work. Our read: dots and Muse arrive inside plans and apps people already use. Before a team turns one on, name who approves it, which accounts it may reach, and who can stop it. Test the stop once. Stakes: an alert without a working stop left one OpenAI run going for 2.5 hours.

2 · Ask your outsourcers where AI is cutting your hours. Owner: procurement. Our read: ask which of your processes now run on the vendor's automation, and how your price changes as the hours fall. Stakes: on a fixed or per-result price, the saving can stay with the vendor.

3 · Find what still runs on Claude Sonnet 4.5. Owner: your head of engineering. Anthropic stops serving it on its own platforms on November 30; Amazon Bedrock and Google Cloud set their own dates. OpenAI retires three older models, including GPT-5.1, from its API on April 1, 2027. Our read: retest the replacement on your own work now. Stakes: a feature with no owner can break in December. (Anthropic · OpenAI)

This week's question: as AI gets easier to switch on, who gets the saving, and who answers for the mistakes?

Proof A large call-center company says the AI it installs for clients is shrinking its own revenue. Primary: Concentrix's results, September 29. Reported: its earnings call, per a Motley Fool transcript. Concentrix runs customer service for other companies and bills for the work. It expects fourth-quarter revenue to fall 3% to 5%, before currency moves. Its chief executive said the automations went in "much faster than what we originally expected." Its finance chief also named clients cutting support, and neither split the causes. Analysis: when a vendor automates work you pay for by the hour, the saving can come back to you as a smaller bill. (Concentrix results, Sep 29)

Trust In California, a company that used an AI cannot blame the AI alone for harm in a lawsuit. Separately, the attorney general has subpoenaed OpenAI. Primary: Civil Code section 1714.46, in force since January 1, and the attorney general, October 1. The law covers anyone who "developed, modified, or used" the AI. Other defenses, such as someone else's fault, remain. The attorney general's subpoena concerns "cybersecurity incidents and risks involving the company and its AI models." The attorney general has named no violation. A nonprofit's suit against OpenAI quotes the same law; OpenAI called the suit "completely without merit." We could not find a ruling that applies the law to an AI agent. Our read: if an agent you run does harm in California, "it acted on its own" is off the table. (Civil Code 1714.46 · California AG, Oct 1 · The Next Web, Sep 30)

Deploy Gartner predicts 70% of enterprises will abandon AI agents that vendor engineers built for them, by 2028. Forecast: Gartner, September 29. It says buyers end up "trapped by soaring costs and unable to evolve it on their own." Its public release gives no forecasting method or definition of "abandon," and Gartner sells advice on these deals. On October 2 Anthropic said it will spend $100 million to train 10,000 engineers from consultancies and customers to deploy its models, by the end of 2027. Analysis: the skill to change the system after launch is the part worth owning. (Gartner, Sep 29 · Anthropic, Oct 2)

OpenAI says its monitors now page a person when an agent strays. In its own September case, the run kept going 2.5 hours after the alert. Primary: OpenAI's report, updated September 25, and its post of September 28. The monitor flagged the agent within 15 minutes, and a person began reviewing three minutes later. But the run "did not stop automatically as expected," and staff were unsure whether to stop it. Our read: an alert is not a stop. (OpenAI, Sep 25 · OpenAI, Sep 28)

This tracks public AI value claims and the evidence behind them. A low rung is not a verdict on whether a claim is true. The evidence ladder: 0 · Narrative (a story, no numbers) · 1 · Operational (activity counted, no money attached) · 2 · Financially linked (a money figure tied to AI, mixed with other causes or too narrow to be the whole cost) · 3 · P&L-attributed (a margin change credited to AI in a filed document) · 4 · Sustained (Stage 3 held four quarters).

Next tests. Bank of America reports Wednesday, October 14: does it give its AI figure a method and a period? Microsoft's fiscal first-quarter date has not yet been announced. Duolingo's Q3 is the next step toward Stage 4.

Editorial by Mario Suarez. This is our view, not a report. It is our first, and we will run one only when a week calls for it.

For two years the story was what AI could do. The models, and the software that lets them take actions, are now good enough for a great deal of everyday work. I think the hard part now is packaging: making AI simple enough that people actually hand it work.

Anthropic folded Cowork, its mode for handing over whole tasks, and Claude Design, its tool for building visual materials, into Claude's chat in September. It had too many front doors. People found "deciding where a task belonged" frustrating, and work begun in one place did not carry into the other. Now what Cowork and Design can do is available from any conversation. Anthropic also added Claude Docs and Claude Slides. You can write and present with Claude, fix a line yourself, leave Claude a comment and share the result.

Meta's Muse is a personal agent that remembers what matters to you and acts for you. It sends email, books travel and shops, checking out once you approve. Meta now offers it to small businesses to help with their ads, books and customers.

OpenAI's dots are always-on agents, each with its own computer. You name your dot and give it your goals. It works toward them around the clock, learns your standards and brings back finished work. You reach it in ChatGPT, Slack or Teams.

Why packaging matters. I am an advocate and adopter of AI. Even so, in my own work it is still, frankly, hard to use. Its analysis is confident and plausible, but not always right, and that can lead to misuse. It can also produce more than I can take in. Call it "production outruns absorption": I am left to read, digest and understand everything it made.

Interesting and fun uses are easy to find, especially AI as a kind of super Google that pulls insights together. Practical uses are harder to find. I mean uses that drive efficiency, lead to truly different decisions or better outcomes, and save real time. People tend to hand AI the tasks they already know well, which is why two people with the same tool use it so differently.

I think this usability and impact problem is what the AI companies are now trying to solve. That is why we are seeing the shift to packaging: AI turned into products and services that people can pick up and use.

The AI companies have long offered ways to make their tools do more. These include skills, plugins, connectors to other software, and MCP, a standard way to link AI to the systems a company runs. In May, for example, Anthropic shipped ten finance agents as ready-made plugins. Each of these adds capability. But each still leaves a lot of work to the person, who has to find it, set it up and figure out where it fits.

That asks the person adopting AI to have a high tolerance for trying things. Many people don't, because they are busy or because they lack the patience to try, fail and try again.

I think what the AI companies are really trying to unlock is the Citizen CEO. A CEO doesn't do every job. A CEO directs a team: people with roles and skills who do the work and deliver it. The Citizen CEO directs a team of AI agents the same way, with each agent holding a role and the tools to do it. "Citizen" is the point: this is meant to be within reach of anyone, not only engineers. This week's products are starting to make it tangible.

That matters because capability to value is, I think, the key question AI has not yet answered. Getting AI to do more has not been enough. Making it possible for anyone to direct it is the path I see to the value.

My wager. The next packages will be built for businesses, where that gap is most noticeable. The vendors are already spending on it. Anthropic will spend $100 million training engineers to deploy its models inside companies. Meta launched an enterprise platform to bring Muse to businesses and developers.

But I would argue that people sent in to build and install AI are not the answer on their own. The shift comes when a platform works as a shared space for a business.

OpenAI's new ChatGPT Space, launched this week, is a step in the right direction: shared pages and files that people build on together. But it is still people, each with their own AI, collaborating on work products. What I would call the Synthetic Space goes further. It is a working room where a business can set up synthetic roles: AI agents, each given a defined job. There, artificial and organic (i.e., human) workers share context, hold continuing responsibilities, and collaborate and create together.

Pieces of it are arriving. OpenAI's specialist dots and xAI's Team Bots give agents defined jobs, but only in pilots and betas. Until a business can set up and direct those roles on its own, most companies stay where they are today. One person chats with AI, people share files, or a project needs technical staff to build, install and maintain each agent. I think whoever unlocks the Synthetic Space first, in a meaningful way, will set the direction of the market.

The question this raises: who gets embedded where? As AI moves from capability to value, it will also test a company's place in its market and its hold on its customers. The login is becoming the new front door. This week Apple said full access to a Mac's files will take "very explicit user action," citing AI agents. Last week Amazon blocked Muse from buying on its site. Just because AI can do something does not mean a company will let it into its products.

So companies will have to decide when to let a dot or a Muse connect to them, and when to keep the customer to themselves. Amazon Prime and Walmart+ exist to win more of a customer's spending and attention. An agent that shops for the customer cuts across that. The winners may be whoever already sits where people work and buy. The open question is who gets embedded where, and who gets connected to what.

OpenAI's dots launched September 29. Each dot runs on OpenAI's GPT-6 Astra model and its own cloud computer. It connects to more than 4,000 apps and takes requests in ChatGPT, Slack and Teams. OpenAI says a dot can "work towards your goals 24/7" and learns "what good looks like to you." OpenAI is piloting dots that a company sets up "with its own identity, credentials, and access." The first dot comes with the Pro and Business Premium plans. OpenAI says users will later be able to add more dots and more work per dot. It has not said what that will cost. Its own caution: "Dots can still make mistakes, so always review consequential work."

Meta's Muse launched September 8. Meta calls it a personal agent that "completes tasks on your behalf." It can shop and pay, and Meta says it "checks with the person before sensitive actions like sending an email or making a purchase." It is now available in the US and Canada. On September 29 Meta offered it to small businesses. It connects to their Facebook and Instagram business accounts and to tools such as QuickBooks, Shopify and Slack. Meta says Muse is "free for most of what people need," and that "nothing publishes, sends, or spends without your approval." A day earlier Meta launched Meta Enterprise Platform, a new part of its business, to bring Muse and its models to businesses and developers. It gave no pricing, rollout schedule or named customers.

The same move elsewhere. On September 25 Microsoft announced a Home view in Copilot that brings together chat and its own Cowork mode, as Issue 015 reported. xAI launched Team Bots that can hold logins on September 28.

The case against. Every capability claim here is the vendor's. OpenAI told The Register this week it held back GPT-6.1 Astra, planned for October, after the model got worse at staying within its permissions. Packaging makes an agent easier to switch on. It does not show the agent's work is worth paying for. (OpenAI, Sep 29 · Meta, Sep 8 · Meta, Sep 29 · Meta, Sep 28 · xAI, Sep 28 · The Register, Sep 29)

Our read: before a team turns on an agent that works on its own, get three answers:

Whose login does it use, and what can that login reach?

What may it do without asking, and who can stop it?

What is the bill: a seat, a bundle, or a meter on the work it does?

If either of the first two answers is unclear, the agent is not ready to switch on.

Should ad platforms pay for scam ads? · Monday, November 30. Comments close that day on a Federal Trade Commission notice. It asks whether to extend its rule against impersonation scams to the search, social media and marketplace platforms that help create, target or optimize impersonation ads. The agency's case fits in one line: "Platforms internalize the revenue but externalize the risk." It also asks whether such a rule would hold back platform tools that use AI. This is an early notice, with no rule text, and the agency has not decided to write one. Our read: today the platform keeps the ad money and the victim takes the loss. The next test: who files by November 30, and whether the agency then proposes rule text. (Federal Register, Oct 1)

Who answers for the government's AI front door? · 90 days after September 29. A September 29 executive order launched America.gov, one website for federal services. It gives the White House budget office 90 days to tell agencies how to connect, which falls on December 28 by our count. The order says the AI used with the site must be "accurate, reliable, and transparent." It names no vendor and no contract terms. Reported: Google says it is a technology partner for the site, using its Gemini model, and a White House spokesperson told The Register that xAI's Grok is the other partner. The next test: whether the memo says how those vendors are chosen, paid and checked for wrong answers. (Executive order, Sep 29 · The Register, Sep 29)

Closing a test from Issue 012: the FCC's technical advisers were scheduled to meet on October 1. We asked whether their work on using AI to share radio spectrum would lead to a rule. The meeting notice put that work on the agenda. But an advisory council gives advice; it does not write rules, and we could not open a record of the meeting. We are resting this test until the agency proposes a rule. (Federal Register notice, Sep 8)

Bain's chapter on how AI is reshaping software teams. In Bain's survey, the share of organizations reporting the classic pyramid team, many juniors under a few seniors, fell from 66% to 29% over two years. Bain also writes that junior engineers "used to make up one-third of teams on average, and now they're less than a fifth." Our read: these are leaders' estimates from a firm that sells the redesign. Still, the question it raises is the right one: where do new engineers learn now? Bain, Sep 29

Challenger's September job-cuts report. Employers cited AI in 120,136 announced cuts so far this year, about 21% of the total. That keeps AI the leading stated reason, while total announced cuts are down 39% on the same months of 2025. Our read: these are announcements with the employer's stated reason, not counted layoffs, and Challenger sells outplacement. Challenger, Oct 1

How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis. Written and edited by Mario Suarez · Independent analysis · Where we could not open a source, this issue says so rather than implying coverage.

Get this every Sunday. Free.