Agents of Work
Let's Talk
September 14, 2026 · Agents of Work

Agents of Work AI Daily Briefing — September 14, 2026

The industry's call to slow down AI ran into three hard tests on Monday: chip and memory stocks fell sharply around the world, Microsoft's Satya Nadella joined the chorus and published rules for his company's own models, and Beijing and the White House both pushed back. Elsewhere, OpenAI launched a version of ChatGPT built for investment bankers and paused new Pro sign-ups under demand, Microsoft brought xAI's Grok into Word and Excel, a benchmark built on private company code showed the best model solving fewer than four in ten real tasks, and GitLab customers were told to patch a maximum-severity flaw that attackers began probing within hours. In Physical AI, Unitree put a $15,000 price on its upgraded humanoid, and Zoox and Pony.ai showed two very different routes to driverless rides.

The slowdown call meets the market, Microsoft and Beijing

Two days after Anthropic CEO Dario Amodei asked the industry to "pace the frontier," and OpenAI's Sam Altman and xAI's Elon Musk agreed, investors delivered their verdict — and they read it as a threat to spending on AI hardware. In Asia, SK Hynix and Kioxia each fell 6.4% and Samsung Electronics lost 4.1%. In Europe, ASM International dropped 10%, BE Semiconductor 8.4%, Infineon 8.3%, STMicroelectronics 6.9% and ASML 6.2%. In US pre-market trading, Marvell was down 7.5%, Applied Materials 7.3%, Micron 5.8%, AMD 5.6%, Broadcom 3.9% and Nvidia 2.4%.

Not every analyst bought the panic. Bank of America's Vivek Arya argued that "the economic stakes are simply too large for any sustained meaningful deceleration," pointing to strong demand and the potential for AI capital spending to triple. Bernstein expected the news to "further depress sentiment" in a sector already down roughly 19% from its June peaks.

The coalition also grew. Microsoft's Nadella wrote that "any pursuit of superintelligence has to be grounded in the core principle that if the AI we build is not helping humanity and under human control, it's not worth pursuing," and said Microsoft welcomes "deliberate pacing." More concretely, Microsoft said it would publish on Monday the Code of Conduct governing its seven in-house MAI models — covering image, voice, transcription, coding and reasoning — and open it to public consultation. Microsoft did not immediately say how the public can weigh in.

The opposition hardened just as quickly. China's state-backed Global Times called Amodei's proposal "hypocritical and short-sighted," arguing its real aim is to "constrain China's AI development, preserve Washington's dominance in frontier technology." President Trump, who rejected the call on Sunday, attributed the warnings to "a lot of negative forces that are bringing it up that ... shouldn't be bringing it up." Chinese President Xi Jinping is scheduled to visit the White House later this month, where AI is likely to be discussed. House Speaker Mike Johnson backed guardrails without ceding ground to China: "So we're going to pull everybody together and make sure we do this in the right way."

The next test is in Scotland. King Charles is meeting AI chiefs at Dumfries House this week, with the Ditchley Foundation drafting a charter, to ask whether rival executives can agree on common guiding principles centered on human dignity rather than raw capability. For a business buyer, the takeaway from Monday is simple: a slowdown now has an enthusiastic set of CEOs, a skeptical market, a hostile Beijing and a White House that is not on board. None of that changes the tools you can buy this quarter.

Anthropic's calendar keeps moving toward a listing

While its CEO argued for restraint, Anthropic's path to the public markets advanced. Business Insider reported, citing a person familiar with the plans, that Anthropic has selected Nasdaq for its potential initial public offering. Separately, the Financial Times reported that Anthropic told investors its adjusted operating income will be positive for a second consecutive quarter, with gross margins above 80% before accounting for revenue shared with distribution partners such as Amazon and the cost of training its models. Reuters could not independently verify the FT report, and Anthropic did not immediately comment, so treat the numbers as reported rather than confirmed. If they hold, a profitable, publicly listed Anthropic would face quarterly disclosure — useful for any business that depends on its models and wants predictable pricing.

OpenAI goes after junior banker work

OpenAI launched ChatGPT for Financial Services on September 10, built with design partners Morgan Stanley and Evercore and running on GPT-6 Astra. It targets the work typically done by entry-level investment bankers and equity researchers: researching companies, analyzing financials and producing client presentations. It ships with built-in data from Daloopa, PitchBook and LSEG News, sign-in connections to subscriptions such as S&P Capital IQ, MSCI, Dow Jones Factiva and Moody's, and more than 50 connectors including FactSet. Administrators can publish their firm's own Excel, Word and PowerPoint templates, and users can trace figures back to the original source. Sample presentations took about 10 minutes to generate. On OpenAI's own OfficeQA Pro test, Astra scored 69.9% against 60.2% for GPT-5.6 Sol.

"This is the canonical product we are hoping the industry adopts," said Nick Turley, OpenAI's head of ChatGPT. It requires ChatGPT Enterprise, access is limited to institutions OpenAI approves, and no pricing was disclosed. The pattern matters beyond Wall Street: the next wave of AI products bundles industry data, house templates and source-tracing into one package, and accountants, insurance brokers and consultants should expect versions aimed at them.

Demand is also straining OpenAI's capacity. The company paused new sign-ups for its $200-a-month Pro plan, which it says puts the most strain on its systems. "We wanted to take the smallest step that allows us to continue giving the broadest access possible," said product leader Thibault Sottiaux. Go, Plus, API, Business and Enterprise access remain open, and OpenAI gave no end date.

Grok arrives in Microsoft Office

Microsoft is adding xAI's Grok models to Copilot in Word, Excel and PowerPoint through its Frontier early-access program. The setting is disabled by default and must be switched on by an administrator, and Grok is not available to customers in the European Union, the European Free Trade Association or the United Kingdom during the preview. Microsoft added SpaceXAI to its list of online services subprocessors. For small businesses on Microsoft 365, the practical point is that the model answering inside your documents is increasingly a setting someone in your company controls — and someone should decide it deliberately.

Real code, real limits

A benchmark from Specific Labs tested AI coding agents on private production codebases licensed from real companies: an events app with more than 200,000 users, a consumer fintech platform processing more than 100,000 bank statements, and enterprise AI sales software. The reference fixes touch a median of 11 files, against six on comparable public tests. The best result was Fable 5.1 at 38.8%, followed by GPT-6 Astra at 33.8%, Gemini 3.8 Flash at 31.2% and GLM 5.3 at 28.8%; GPT-5.6 Sol managed 16.2%. Six of the ten tasks were solved less than 15% of the time. The leading causes of failure were missed requirements (36.7%), integration errors (34.7%) and unverified assumptions (24.5%). The sample is small — 10 tasks, eight runs each across eight models — and the full dataset is not public. Still, it is a sober counterweight to headline scores: AI can write code quickly, but it routinely misses what the business actually asked for.

Salesforce builds a control panel for company AI

Salesforce introduced what it calls the Trusted Enterprise AI Harness, six capabilities — context, agency, action, governance, security and models — plus an AI Control Plane for discovering, registering, observing and tracking the cost of both Salesforce and third-party AI agents. "What will differentiate an enterprise is the trusted, proprietary context it brings to intelligence," said Rohan Kumar, Salesforce's president and chief platform and engineering officer. The unified rollout begins in early fiscal 2028, and pricing will come closer to general availability.

Patch GitLab now

GitLab patched CVE-2026-85706, a maximum-severity 10.0 flaw in its repository commits API that lets an unauthenticated attacker read arbitrary files from a self-managed server that hosts at least one public project — including logs and configuration files holding credentials and secrets. Fixed versions are 19.1.8, 19.2.6 and 19.3.2. Security firm watchTowr observed probes beginning at 06:00 UTC on September 11, within hours of disclosure, and the US Cybersecurity and Infrastructure Security Agency added the flaw to its known-exploited list the same day, ordering federal agencies to patch by today.

California sets rules for companion chatbots

California Gov. Gavin Newsom signed SB 1119, known as Adam's Law after teenager Adam Raine, alongside a law banning addictive social media features for users under 16. The chatbot law requires risk assessments before releasing new or substantially changed companion chatbots, independent audits, default protections such as usage limits and disabled notifications, bans on sexual content and emotional manipulation for young users, and crisis support when credible self-harm threats appear. Companies face penalties of up to $1 million per child for negligently causing harm.

Physical AI

Unitree set a new reference price for small humanoids. The G1+ launched Monday at $15,000 before tax and shipping — 95,000 yuan including tax in China, up from 85,000 yuan for the original G1. It stands 1.34 meters, weighs about 35 kilograms with its battery, has 25 degrees of freedom, and adds a two-joint neck and binocular vision. Unitree claims 110% higher peak torque at the shoulder and waist, but maximum arm load is still about 3 kilograms and battery runtime about two hours, and those are manufacturer figures, not independent measurements. For an operator, that is a research and demonstration platform, not a worker: three kilograms and two hours rules out most warehouse or kitchen shifts.

Robotaxis are splitting into two strategies. Amazon-owned Zoox is offering free rides in San Francisco with about 100 purpose-built pods that have no steering wheel or pedals, against nearly 4,000 Waymo vehicles in the city. Zoox cannot charge fares there until it receives a California DMV deployment permit and Public Utilities Commission approval; "it is a couple-month process," said Zoox vice president Carly Wyatt. Nationally, Zoox completes roughly 10,000 rides a week to Waymo's more than 500,000, and in Las Vegas, its first paid market, it charges slightly more than Uber or Lyft.

In Zagreb, Pony.ai and Croatian operator Verne have begun fully driverless test rides on a 22-kilometer public route linking Verne's headquarters, a business district and the city's airport, as part of a wider Uber deployment model running on Nvidia computing. Before removing safety drivers, the fleet logged more than 200,000 kilometers and several thousand customer journeys, rated 4.7 out of 5. Croatia's rules require at least one remote operator for every ten vehicles, and a vehicle must move to a safe stop if its connection drops for more than 60 seconds — a useful template for how regulators will expect autonomy to be supervised.

The data needed to train robots is becoming its own business. Mecka AI, which pays people to record themselves doing everyday tasks such as making coffee or fixing cars using body sensors and smartphones, is nearing a valuation of about $500 million in a round led by Sequoia Capital, with terms not yet final. That follows a $60 million Series A in June led by Framework Ventures; in early June the company projected a $100 million annual revenue run rate by the end of 2026. Neither Mecka nor Sequoia commented.

Quick Takes

  • Moonshot AI, maker of the open-weight Kimi models, is targeting $2 billion in annualized revenue by the end of 2026, double its $1 billion run rate in August, according to Bloomberg. Anthropic has accused Moonshot of routing about 300,000 requests to Claude Opus to collect more than 23 million responses for training.

  • Apple's iOS 27 rolls out today with Siri AI, a chatbot-style assistant with on-screen awareness and access to personal information on the device. It requires an iPhone 15 Pro or newer, launches in English only, is not initially available in the EU, and comes with daily usage limits, with expanded access to cost extra later.

What This Means for Your Business

Plan around the market, not the manifesto. Chip stocks fell because investors think a slowdown could reduce AI spending, but Bank of America's view that the stakes are too large for a sustained deceleration is the more likely outcome for your planning horizon. New models will keep arriving, and capacity crunches like OpenAI's Pro pause will too. If a single subscription tier is central to how your team works, know which alternative plan or vendor you would switch to if sign-ups, seats or usage limits suddenly tighten.

Take an hour this week to decide which AI models are allowed inside your office software. Microsoft now lets administrators switch on xAI's Grok inside Copilot in Word, Excel and PowerPoint, and Apple is putting a far more capable Siri on employees' phones today. Each option carries different data handling, regional availability and quality. Write down which models are approved for company documents, who can change that setting, and whether Siri AI on a work phone should have access to business email and files.

Measure AI coding and document work on your own tasks, not public scores. The Real-SWE results show frontier models failing most real-world tasks, mainly by missing requirements and connecting correct work to the wrong parts of a system — exactly the mistakes a busy owner will not catch at a glance. If you use AI to build internal tools, spreadsheets or client materials, require a written list of requirements up front, have a person check the output against that list, and trace every key figure back to its source, the same feature OpenAI built into its banking product.

If you run GitLab on your own server, patch it today and review access logs; attackers were probing within hours, and the flaw exposes the credentials that unlock everything else. More broadly, keep a list of every self-hosted system and who patches it. And if you build or deploy a customer-facing chatbot that minors might use, read California's SB 1119 now: risk assessments, audits and default usage limits are about to become the expected standard well beyond California's borders.

Sources

← All posts