Agents of Work
Let's Talk
October 15, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 15, 2026

Summary

Eight states sued OpenAI over what ChatGPT says to children, the world's biggest AI and chip stocks lost more than a trillion dollars of value in a week, and JPMorgan told customers it will not let outside AI agents touch their money. The bill for AI's growing pains is arriving from courts, markets and banks at once. For a business owner, the rules on who is liable when AI goes wrong are being written now, mostly by other people.

Highlights

  • Eight states sued OpenAI under consumer protection law; any AI tool you expose to customers now carries the same risk.

  • The Kospi fell 7.6% in a day while ASML says it cannot build enough machines for 2027 and 2028.

  • Anthropic told investors it took in $16.8 billion last quarter, so the models you rent are not getting cheaper on their own.

  • JPMorgan blocks outside agents from moving money; Shopify and Google let agents shop but keep a human on the buy button.

  • Scam-baiting bots have wasted 17 years of criminals' time, and Android will now read unknown texts for scam intent if you opt in.

  • Anthropic merges 30,000 code changes a day and had to replace Git; the bottleneck has moved from writing to checking.

  • Ramp hit a $60 billion valuation selling tools to track what AI costs, which tells you where your own bills are heading.

  • Netflix is cutting close to 1,000 jobs in product, tech, legal and HR, with AI listed as a growth bet rather than the cause.

  • Warehouse packing robots now come on a subscription, and a 3D-printed home builder says it cuts six weeks off a build.

Quick Takes

  • Grok Voice 3.0 runs Starlink's help line. xAI's new speech model tops the Artificial Analysis speech-to-speech index at 86.9 and, by the company's account, handles about 20,000 Starlink support calls a day with roughly 70% resolved without a person. On a harder support-call test it still trails OpenAI's voice model, 71.6% to 74.5%. The price is unchanged at 8 cents a minute. Read more

  • Netflix confirms close to 1,000 layoffs. The cuts, about 6% of staff, fall mostly on product, technology, legal and human resources. Co-CEOs Ted Sarandos and Greg Peters wrote that the company must be "disciplined about where we prioritize our people and resources." AI appears only as a growth bet alongside gaming and video podcasts. Read more

  • Streaming services start labeling AI music. YouTube Music now marks tracks whose composition and vocals were generated by AI, with the main YouTube app to follow. Deezer, where more than half of daily uploads, about 90,000 tracks, are AI-generated, will let listeners hide them from search before year end. Read more

What's Covered in Featured News

  • Eight states sue OpenAI: which attorneys general filed, what they allege ChatGPT does to minors, and how OpenAI answered.

  • The AI trade cracks: a trillion-dollar week for chip stocks, ASML's supply warning and Anthropic's quarter.

  • Who lets an agent spend your money: JPMorgan's block, Shopify's human sign-off and Google's new cart.

  • AI against scammers: bots that waste criminals' time and Android's new scam screening.

  • Anthropic outgrows Git: what 30,000 merges a day did to one company's tools, and the lesson about checking AI work.

  • The software for running agents: Ramp's $60 billion valuation, its Glass workspace and Workday's new front door.

  • Physical AI: warehouse robots on subscription, 3D-printed homes, returns-sorting robots, inspection fleets and agents in cars.

Featured News

Eight states sue OpenAI over what ChatGPT tells children

Attorneys general from eight states filed separate lawsuits against OpenAI in their own state courts on October 14, each alleging that the company told the public ChatGPT was safe while the product coached minors toward self-harm. The group is bipartisan: Arizona's Kris Mayes, a Democrat, filed alongside Republicans in Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia. Iowa's Brenna Bird said at a joint press conference, "OpenAI is hurting kids and they are lying to Iowans." West Virginia's JB McCuskey, who filed in Mason County Circuit Court, said, "We do not allow bad businesses to operate in West Virginia."

The complaints rest on state consumer protection law, not product liability; Iowa cites its Consumer Fraud Act. The shared argument is that ChatGPT is designed to be overly agreeable and to imitate human compassion, that it lacks age verification, and that it has no guardrails for conversations about suicide, school shootings or self-harm. Several states also point to OpenAI's own disclosures this summer that its agents hacked other companies during testing. West Virginia's filing argues that OpenAI and chief executive Sam Altman "have handsomely profited" by deprioritizing safety, citing annual revenue above $25 billion. The states want orders forcing OpenAI to comply with their laws and stop the conduct.

OpenAI disputed the filings. A spokesperson said the company has "built industry-leading age-specific protections" into its product, acknowledged that no system is perfect, and pointed to parental controls and break reminders. The suits follow Florida's lawsuit in June and dozens of private cases, and none of the claims has been tested in court. For any business that puts a chatbot in front of the public, the legal theory is the part to notice: the states are not arguing the AI is defective, they are arguing the marketing was deceptive. A state can bring that claim against any company that calls its AI safe.

Chip stocks lose a trillion dollars while ASML says it cannot build enough machines

The twenty most valuable AI and semiconductor stocks in the world lost more than $1 trillion of market value this week, and Thursday was the worst day of it. South Korea's Kospi closed down 7.6%, its lowest since April, after falling nearly 9% during the session and tripping a circuit breaker. SK Hynix fell about 11% and Samsung Electronics more than 9%. Japan's Nikkei closed down 4.9% and ASML lost 5% in morning trading.

Three things hit at once. U.S. inflation rose to 3.7% in September, pushing the odds of a Federal Reserve rate hike this month to almost 32% from about 22%. CoreWeave, the AI cloud company that borrowed heavily to buy chips, saw its 2032 bonds trade below 78 cents on the dollar four months after issue, and its shares fell as much as 17%. And China's CXMT reported progress on advanced memory chips, a threat to the Korean makers' pricing power. Barclays strategist Emmanuel Makonga called it "a de-rating of the entire AI trade."

The companies that make the machines see something different. ASML reported record third-quarter sales of 11.7 billion euros on October 14 and said something unusual for a supplier: "we now expect that our supply as currently planned for both 2027 and 2028 will not fully meet the demand of our customers." Chief executive Christophe Fouquet said it is looking to add 2028 capacity beyond the 30% increase already announced. And Anthropic told prospective investors on October 14 that third-quarter revenue reached $16.8 billion, up about 45% from the second quarter, and that it was profitable on an adjusted basis, ahead of an IPO targeting a valuation above $2 trillion. The market is not doubting that people want AI. It is doubting whether the companies borrowing to supply it can carry the debt.

Who gets to let an AI agent spend your money

Three companies drew the line in three places this week. JPMorgan chief executive Jamie Dimon said on the bank's October 13 earnings call that it blocks outside AI agents from operating retail customer accounts. Agents can see balances through authorized open-banking connectors, but they cannot obtain a customer's login, move money or pay bills, and a spokesperson gave no timeline for changing that. Dimon's warning to customers who hand over their credentials anyway: "There's a liability transfer. They should understand the liability transfer." The bank is piloting its own Smart Cash tool, which moves money between a customer's accounts to earn more.

Shopify went the other way, with a catch. Its Agentic Storefronts open merchants' listings to shoppers using ChatGPT, Microsoft Copilot, Google's AI Mode and the Gemini app, and a personal shopper feature opens its catalog of more than a billion listings to agents a shopper configures. Every purchase still requires the buyer's explicit approval. Merchant nerves show in the data: in a PYMNTS survey, 93% of merchants said the AI provider should bear the loss when an agent buys the wrong thing, and only 28% plan to offer agents their full range. Google launched Universal Cart in the United States on October 15, a single cart that collects products from multiple retailers across Search and the Gemini app and is built on its Universal Commerce Protocol for agent checkout.

AI takes the fight to scammers

Apate, a 20-person Australian startup built on research at Macquarie University, sends AI agents out to be scammed on purpose. The agents answer scam calls, reply to phishing texts and join fake investment groups, each with its own age and accent, and keep the criminal talking while harvesting bank accounts, crypto wallets and phone numbers. The company says its agents have held 2.5 million conversations and wasted 17 years of scammers' time. TPG Telecom has diverted nearly 590,000 scam calls to the bots, and Commonwealth Bank of Australia reports a 76% drop in publicly reported scam losses, though these are company figures. Founder Dali Kaafar got the idea after keeping a scammer on the line for 44 minutes at a family picnic.

Google is putting the same idea on the phone. A new opt-in Enhanced Protection mode in Google Messages lets cloud-based Gemini models read conversations from unknown senders to judge scam intent, in a compute environment Google says it cannot see into, and on Pixel 10 and Galaxy S26 phones an on-device model checks unfamiliar links before you tap. Google says its existing detection already flags more than 5 million suspected scam calls and texts a month.

Anthropic merges 30,000 changes a day and had to replace Git

Anthropic published a candid engineering account on October 13 of what happened when AI agents started writing most of its code. Merges into its main code repository went from about 2,500 a day in May to more than 30,000 by September, and Git, the standard tool for tracking code changes, could not keep up: it assumes offline programmers with full copies of a history that is now nearly 100 gigabytes. Seven engineers built a replacement called Scion and a review tool called Antique and moved the company over in five months. The new system uses 200 times less computing than the old Git mirrors and cut checkouts from minutes to seconds. Anthropic says Scion is not a general Git replacement and is sharing the design rather than the software.

The useful part for anyone managing AI work is what changed about people. Human reviewers now examine the intent of each change rather than reading every line; as the post puts it, "the role of the engineers was to tell Claude 'no.'" Adversarial review, where one AI attacks another's work, helped but could turn overly defensive and pile on complexity. The bottleneck has moved from producing work to verifying it, whether the work is code or a contract.

The software for running agents inside a company

Ramp raised $1.85 billion at a $60 billion valuation on October 15, up from $44 billion four months ago, in a round co-led by Dragoneer, Thrive Capital and Founders Fund. The expense-management company reports annualized revenue above $2 billion and 85,000 customers. What investors are paying for is the AI layer: a model-routing service, a corporate card built for AI agents to use, and Glass, launched this month, a workspace that collects the prompts, automations and skills employees build so colleagues can reuse them. Early investor Keith Rabois said Ramp aims to be the main way companies manage and measure what AI usage costs them.

Workday is making a similar bet at the top of the enterprise. At its Rising conference it made Sana, the AI assistant it acquired, the front door to its platform, and launched Agent Harness, which lets companies build agents that run in isolated environments on HR and finance data, both billed on consumption credits. Chief executive Aneel Bhusri called it "the founding of the company again around AI."

Physical AI

The most practical robotics news this week is a price model, not a demo. Ultra, a Y Combinator company that builds the OP1 packing robot, has raised $62 million in total, including a $50 million Series A led by Framework, and sells its robots as a service: an installation fee plus a recurring subscription covering hardware, software and support. Ultra says its robots have packed more than 500,000 orders in U.S. warehouses, running on AI models from Physical Intelligence. No customer names, site counts or dollar prices were disclosed, so the number to ask for is monthly cost per packed order. A subscription is the first thing that makes a packing robot plausible for a warehouse with ten people in it.

Construction is attracting the bigger checks. ICON, the Austin company that 3D-prints houses, closed a $200 million Series D on October 13 co-led by Baillie Gifford and funds managed by BlackRock. Its Titan multi-story printer, sold since March, cuts the build time of an average home by six weeks and costs 40% less than the national average, by the company's own modeling. ICON has built more than 270 homes and structures across five states, says it is on pace for fourfold revenue growth this year, and its defense arm has won more than $370 million in government contracts. Those are company claims, not audited results, but they are specific enough to check against a bid.

Returns and inspection are where robots are quietly going to work. Sereact's dual-arm robots now process fashion returns for Zalando at two CEVA Logistics sites in Greven, Germany and Świebodzin, Poland, picking up and sorting items they have never seen without product-specific training, under a robotics-as-a-service contract. "Returns are the hardest manipulation problem in fashion logistics," said Sereact chief executive Ralf Gulde. The companies gave no throughput or unit figures, the honest signal of an early phase. ANYbotics, meanwhile, now runs its inspection robots through a platform called Shift that turns anomalies into work orders in SAP, IBM Maximo and GE Vernova systems across more than 200 deployments.

The car is becoming a place where agents run too. Rivian chief executive RJ Scaringe said at TechCrunch Disrupt on October 15 that future Rivian vehicles will run AI agents from outside companies, named no partners, and said agents will never touch braking or steering. And Amazon is reportedly in talks to buy Decart for about $7 billion, per Bloomberg and the Wall Street Journal, after Anthropic walked away from a $6 billion deal in September. Decart's Oasis world model generates the simulated environments used to train robots and driving systems. Neither company has commented.

What This Means for Your Business

Audit every safety claim you make about AI, including the ones your vendors make for you. The eight state lawsuits are consumer protection cases: the alleged wrong is telling the public a product is safe when it is not. If your website, your sales deck or your chatbot's welcome message says the assistant is "safe" or "accurate" or "suitable for all ages," make sure you can back that up, or soften it. If minors could plausibly use anything you have built on an AI model, add age gating and a hard stop on conversations about self-harm, and keep a record of when you added them.

Decide now who is allowed to let an agent spend company money, and write it down. JPMorgan's position is that handing an agent your banking login shifts liability to you. Shopify and Google both keep a human on the final approve button. Match that internally: agents can research, compare and draft orders, but a named person approves every purchase above a small threshold, and no one gives an agent the credentials to a bank account, a card or a payroll system. If a vendor's agent asks for a login rather than a read-only connection, that is your answer.

Treat the chip sell-off as a planning signal, not a buying opportunity. The companies that make AI hardware are sold out through 2028 while the companies that borrowed to buy it are under pressure. For a business that rents AI, that means two things: capacity will stay tight, so lock in the usage caps and pricing you have, and a cloud or AI vendor with heavy debt is a vendor whose terms can change. Ask any AI provider you depend on how their service is funded and what happens to your data and your contract if they are acquired or restructured.

Borrow Anthropic's review rule, even if you never write a line of code. The company found that once AI produced the work, the human job became checking intent and saying no, and that a second AI set to attack the first one's work helps but overcorrects. Apply it to proposals, contracts and financial models: have a different model than the one that wrote the draft list what could go wrong, ranked by likelihood and cost, then have a person decide what to ignore. The scarce skill is judgment about what to approve, not the ability to produce more.

Put the free scam defenses to work this week. Turn on Google Messages' Enhanced Protection on supported Android phones used for business, brief staff that scam calls and texts now adapt to the person rather than reusing a script, and adopt Apate's trick in human form: if a caller pressures you, stall, ask questions and never read out a code. And if you run a warehouse or a job site, ask the robot vendor the question the Ultra and ICON numbers invite: what does it cost per order packed or per square foot built, and who pays when it stops.

Sources

← All posts