Agents of Work
Let's Talk
November 4, 2026 · Agents of Work

Agents of Work AI Daily Briefing — November 4, 2026

Summary

AI is moving deeper into the tools businesses already use, and the people responsible for those tools are now writing down what can go wrong. Apple opened Siri AI to every app developer, CrowdStrike shipped an off switch for AI agents, and Apple, Goldman Sachs and Revolut each showed a different way of keeping AI useful without letting it run up costs or risk.

Highlights

  • Any iPhone app can now plug into Siri AI in beta, so customers may soon book and order by voice.

  • Siri AI reaches customers in a few months; App Store submissions open in spring, giving businesses a planning window.

  • Revolut cut its AI coding bill about 90% by running an open model on its own hardware.

  • Most of Revolut's gains came from debugging and ticket work, not from writing new code.

  • CrowdStrike now sells a switch that stops a misbehaving AI agent and blocks it from restarting.

  • A developer says Postman's desktop app quietly installed plugins into Codex and Claude Code; check your agent settings.

  • Apple's annual filing warns AI agents "may behave in unintended or unpredictable ways."

  • Goldman Sachs pulled AI coding tools from traders and bankers and kept them with engineers.

  • Tutor rents warehouse robots from $14 an hour, with remote humans stepping in when a robot gets stuck.

Quick Takes

  • OpenAI brings in the hackers who got into Anthropic. OpenAI is partnering with Cathedral, a security startup founded by four former DOGE staffers, to have its researchers use OpenAI's models to attack OpenAI's own defenses. Cathedral told the Washington Post it got into Anthropic's internal systems in under twelve hours; Anthropic says it took several days, that model weights and Claude conversations were untouched, and that the method was out of scope. OpenAI also says it is more than quadrupling rewards for critical bug reports. Read more

  • DeepSeek reportedly plans to train its next model on Huawei chips. A Chinese financial outlet, citing unnamed industry sources, reports DeepSeek intends to use Huawei's Ascend chips to build DeepSeek V5 from scratch, a job those chips have mostly not done before. The report is unconfirmed. If it works, it weakens the assumption that frontier AI requires Nvidia hardware, which matters for future AI prices and for export policy. Read more

What's Covered in Featured News

  • Siri AI opens to developers: what Apple's developer preview lets apps do, the timeline, and the limits on who can use it.

  • Revolut's self-hosted coding AI: how a fintech cut its AI coding costs and what it gave up.

  • An off switch for AI agents: CrowdStrike's kill switch and acquisitions, and a reported Postman plugin install that shows why it matters.

  • Apple and Goldman Sachs pull back the reins: a new risk warning in Apple's annual filing and Goldman's narrower AI coding access.

  • ChatGPT at 1.2 billion weekly users: OpenAI's latest usage numbers and what they mean for how customers search.

  • Physical AI: Tutor's $14-an-hour warehouse robots, Tesla's driverless Robotaxi in Fort Worth, and Bedrock Robotics' autonomous excavators.

Featured News

Siri AI opens to every app developer

Apple said on November 3 that any developer can now connect an app to Siri AI, the rebuilt version of its assistant. The access comes as a Developer Preview in the 27.3 betas of iOS, iPadOS, macOS, watchOS and visionOS. Apps plug in through App Intents, the framework Apple already uses for Shortcuts, which lets Siri see an app's content and take its actions. Apple's examples are the kind of thing a customer would actually say: check a delivery, book a reservation, place an order. A new feature called Skills lets a developer teach Siri to chain several actions into one request, and Siri can now pull up anything an app has placed in the Spotlight search index, not only items that fit Apple's predefined categories.

Apple is rolling the same abilities out across 15 of its own apps, including Calendar, Mail, Maps, Messages, Notes, Photos, Reminders and Safari. The company says Siri AI reaches customers "in a few months," and App Store submissions for Siri AI features open in the spring. It currently works in English, French, Japanese, Korean, Portuguese and Spanish, with Traditional Chinese, Norwegian, Turkish and Vietnamese coming.

There are limits. Siri AI is unavailable in the EU on iPhone, iPad and Apple Watch while Apple works with regulators, and it is off for users under 13. Server-side features carry daily caps that vary with demand, and Apple says expanded access "will be available for a fee in the future." For a business with an app, this is the first real chance to be the thing Siri does when a customer asks for a booking or an order, and the spring submission window is the deadline to plan around.

Revolut cut its AI coding bill about 90% by hosting its own model

Revolut, the British fintech, says it cut its AI coding costs by about 90% in under a year by making a self-hosted open model its default instead of paying for commercial tools like Claude Code and Codex. David Abitbol, the company's operations and infrastructure lead, described the setup at the AI Engineer conference in New York. Revolut started with GLM-5.1 from China's Z.ai and now runs GLM-5.3 on 56 Nvidia B200 chips. Before the switch, Abitbol said, some engineers were spending $1,000 to $10,000 a month on AI coding with little return, which pushed the company to cap budgets. The self-hosted version cost about a tenth as much.

The tool now reaches roughly 2,000 engineers, and Revolut reports a 60% improvement in coding time and 70% in lead time. Notably, the biggest gains came from debugging and working through Jira and Confluence tickets, not from generating new code. There were trade-offs: GLM handles text only, so a router sends anything involving images to Gemini or OpenAI, and it has not fully replaced the commercial models. Abitbol was also blunt about very long memory windows: "Long context is a scam," he said, arguing that a fraction of the advertised size is what teams actually use. Running the model on its own hardware also sidesteps the privacy worry about sending data to a Chinese provider's servers.

Security vendors race to put an off switch on AI agents

CrowdStrike announced an AI Kill Switch on November 3 at its Fal.Con Europe conference, available immediately in its Falcon Guardian product. It lets administrators stop a misbehaving AI agent, prevent it from restarting and cut off its network access. "Everyone is building AI agents, but they never had the off button, and now they do," CEO George Kurtz told CNBC, adding that enterprise customers wanted CrowdStrike in place before rolling out Anthropic's Claude Cowork.

The same day, CrowdStrike agreed to buy two AI security startups. The Wall Street Journal put the combined price at about $495 million, roughly $400 million for Fiddler AI and $95 million for Dreadnode; CrowdStrike has not disclosed terms. Fiddler monitors AI agents in production, tracing their reasoning, catching drift and made-up answers and tracking spending, and will be sold as Falcon Observe. Dreadnode lets testers describe an attack in plain language and then runs more than 100 known techniques against a company's AI systems. Both deals are expected to close by January 31, 2027.

Why this matters now showed up in a GitHub bug report the same week. A developer reported that Postman's desktop app, version 12.33.4, automatically installed Postman plugins into both OpenAI's Codex and Anthropic's Claude Code without asking, and added a startup instruction telling the coding agent to load Postman's own skill before handling API tasks. The reporter says the installer also marked that instruction as trusted. Postman had not responded in the thread when we checked. Whatever the intent, it is a clean example of software quietly changing what your AI agent does, and a reason to look at the agent settings on any machine that runs one.

Apple and Goldman Sachs pull back the reins

Apple's latest annual report adds a new risk: AI agents. According to CNBC's reading of the filing, Apple warns that agents "can take actions autonomously," including "taking actions that were not directed, instructed or authorized," and "may behave in unintended or unpredictable ways." It also says its "ability to oversee and control third-party artificial intelligence agents" may be limited. Risk factors are written by lawyers, but this one comes from the company that just invited every developer to wire their app into Siri.

Goldman Sachs, meanwhile, has restricted advanced AI coding tools, including Claude Code and Microsoft's GitHub tools, to its engineers, according to Bloomberg, which cited internal guidelines from October. Traders and bankers had been using them for months to track activity and produce content. In July, chief information officer Marco Argenti had praised non-technical staff building prototypes with AI. A Goldman spokesperson said the decision reflected a "variety of factors including productivity and risk management" and that the experimentation phase is over.

ChatGPT reaches 1.2 billion weekly users

OpenAI says ChatGPT grew from 1 billion to 1.2 billion weekly users in 60 days. Adam Fry, who leads the consumer product, gave the figures at an OpenAI dinner in New York, along with a claim that ChatGPT's search-style questions now amount to about a fifth of Google's volume. That number is OpenAI's own and has not been independently checked. Fry said people increasingly use ChatGPT to look up people, places, prices and health questions that need current information, and that voice and dictation together reach 250 million weekly users. For a local business, the practical point is that a growing share of "who should I call" questions now gets answered by a chatbot reading the web, not by a list of blue links.

Physical AI

Tutor Intelligence, the warehouse robotics company spun out of MIT, raised a $75 million Series B led by Andreessen Horowitz, according to the firm's announcement on November 3. The pricing is the story. Tutor rents robots by the shift, starting at $14 an hour, and when a robot gets stuck a remote human "tutor" takes over and guides it through the task; that intervention becomes training data for the rest of the fleet. Because customers pay by the hour, robot time can be booked as an operating expense like labor rather than a capital purchase. Tutor says it has 160 employees, operates in 20 U.S. states, has quadrupled its fleet since last year and counts small family businesses among its customers alongside large factories. It runs two machines: Cassie, which moves pallets and heavy items, and Sonny, a semi-humanoid with two arms for picking and packing. CEO Josh Gruenstein estimates the pair can handle 80 to 90 percent of manufacturing and logistics work, which is his projection rather than a measured result.

Tesla launched its driverless Robotaxi service in Fort Worth on November 2, its seventh U.S. market, covering about 145 square miles across Fort Worth and several neighboring towns. Scale is still small: an independent tracker showed eight Robotaxis available on Tuesday for an area of about 454,000 residents, and 90% of the 12 trips it had logged cost under $16. Riders book through Tesla's Robotaxi app. For businesses that move people, such as hotels, clinics and event venues, the takeaway is that driverless rides are arriving city by city, but at a fleet size that cannot yet be planned around.

Construction is the other place autonomy is turning into paid work. Bedrock Robotics, which retrofits existing excavators and other heavy equipment with self-driving kits, raised a $600 million Series C at a $4.1 billion valuation in late October, co-led by CapitalG and Emergence Capital, bringing its total funding to nearly $1 billion. Bedrock says its fully autonomous excavators began doing paid work on active commercial sites in August, and its customer base has more than doubled since 2025. New deployments include Wagner Construction in Minnesota, Smith Trucking Company in Florida, Native Excavating in Colorado and C.W. Matthews in Georgia. The retrofit model matters for smaller contractors: it works on machines a company already owns instead of requiring a new fleet.

The common thread is the business model, not the hardware. Both Tutor and Bedrock are selling hours of work or upgrades to existing equipment, and both keep humans close by for the cases the machine cannot handle. For an operator, that means the right questions are the hourly rate, how often a person has to step in and who pays for that time, not how impressive the demo looks.

What This Means for Your Business

If you have an app, put Siri AI on your 2027 roadmap now. List the three or four things customers most often do in your app, such as booking, reordering or checking status, and ask your developer what it would take to expose those through App Intents. The App Store window opens in spring and Apple's own apps are already being rebuilt around this, so being the default answer to "book me a table" or "reorder my usual" is a near-term opportunity, not a someday one. Keep the EU gap and the daily usage caps in mind if your customers are in Europe or use the app heavily.

Audit what your AI agents can touch this week. The Postman report is a reminder that other software can quietly change how your coding or office agents behave. Look at the plugins and startup settings in any agent tool your team uses, remove anything nobody installed on purpose, and make sure you know how to shut an agent off. You do not need CrowdStrike's product to do that, but you do need a named person who knows where the off switch is and when to use it.

Do the Revolut math on your own AI spending, even at small scale. You probably will not run 56 Nvidia chips, but the lesson travels: track who uses which AI tools, what each costs per month and what it actually saves. Revolut found its biggest wins in debugging and ticket work, not writing new code, so measure the boring tasks first. If one or two people account for most of the bill, a cheaper model or a hosted open model may cover most of their work.

Decide who on your team gets which AI tools, on purpose. Goldman Sachs did not ban AI; it matched powerful coding tools to the people who can use them safely. A simple written policy, with general assistants for everyone and code-writing or system-access tools only for people who understand the systems, is cheap insurance.

Finally, check how your business shows up when someone asks ChatGPT for a recommendation in your category and town. If the answer is wrong or missing, fix your website, Google Business Profile and review listings, because those are the pages these tools read. And if you run a warehouse or light manufacturing, ask Tutor or a similar provider for an hourly quote and their intervention rate; it is now a line item you can compare against overtime.

Sources

← All posts