Agents of Work
Let's Talk
October 19, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 19, 2026

Summary

Charles Schwab and JPMorgan are turning away AI agents that try to move client money, while Stripe says the number of agents spending through its wallet has more than doubled in a month. The people who hold the money and the people who build the agents are negotiating who is liable when software buys something, and no law has settled it. The same weekend Anthropic set aside 5% of itself for people AI displaces, and chip stocks posted their longest losing streak since 2008.

Highlights

  • Schwab blocks dozens of AI-agent attempts to move client money every week; agents may read your accounts, not act on them.

  • Stripe counts about 200 agents buying through its wallet, so your customers' software is already shopping; make sure it can check out.

  • Anthropic pledged 5% of its shares to a public benefit fund for AI's labor fallout, with the governance still to come.

  • Chip stocks fell eight sessions straight and Microsoft walked away from a UK data center; the buildout is being repriced, not cancelled.

  • Microsoft reportedly pays Anthropic over $2 billion a year for Copilot, and routes cheap work to OpenAI; your vendor does the same with you.

  • China wants a license before anyone ships the weights of its biggest models abroad, so plan for Chinese open models to stay mid-sized.

  • OpenCode leaked customers' OpenAI, Anthropic and Google keys through an old system; rotate keys you gave any third-party tool.

  • Google DeepMind says most of its code is AI-written, but nearly nine in ten changes still need a human steering.

  • Cybersecurity job postings hit a four-year low while AI security engineer roles tripled; hire for the second, outsource the first.

  • A $7 million forestry robotics round moved 100,000 tons of timber in six months, which is what a real robot deployment looks like.

Quick Takes

  • Workday opens its live data to agents, with a gate. Workday Data Cloud lets AI agents query the HR and finance system of record directly rather than from copies. Agent Lake and Live Data Query are generally available; governed agent actions are early access. Every tool call passes through a gateway that blocks or escalates anything outside policy, such as a spending limit, to a person. Read more

  • The AI phone is here, and the labs want bigger ones. Qualcomm chief executive Cristiano Amon says leading AI companies have told him they need a model of at least 100 billion parameters running almost constantly on a phone by 2028. StepFun's STEPX Neo, on sale in China from October 16 at 3,999 yuan, about $560, builds an agent into the operating system that works across nearly 100 partner apps, with a year of usage included and no price for year two. Read more

  • Illinois convenes an AI Cabinet. Governor JB Pritzker's executive order creates an unpaid advisory group to draft policy through 2027, including whether data center incentives should depend on safety standards and whether existing law can hold AI companies accountable for harm. Illinois already passed the first state law requiring annual AI audits, effective 2028. Read more

What's Covered in Featured News

  • Banks versus agents: why Schwab and JPMorgan are blocking AI agents from moving money, what Stripe's agent numbers show, and the laws that have stalled.

  • Anthropic's public benefit fund: what the 5% pledge covers, what it may and may not pay for, and what is still undecided.

  • The buildout repriced: the chip selloff, Microsoft's exit from a UK data center, and the nuclear plants restarting for AI.

  • Who buys whom: Microsoft's reported Claude deal and Disney's hiring of Character.AI's staff.

  • China's export controls: the draft rule that would license the weights of its largest models.

  • Security: OpenCode's leaked API keys and a model that reasons about its own training.

  • Work: Google DeepMind's AI-written code, and where cybersecurity jobs went.

  • Physical AI: OpenAI's humanoid, Rivian's autonomy ambitions, forestry robots, Honeywell's building agents and Nvidia's agent CPU.

Featured News

Banks start blocking AI agents, while Stripe says the agents keep spending

Charles Schwab chief executive Rick Wurster told Bloomberg TV on October 16 that his firm is actively blocking unauthorized AI agents, including Meta's Muse, from executing transactions for clients. He said a weekly report shows Schwab turning away "tens and twenties" of attempts by major AI agents to reach and move client funds. His reasoning was about liability, not capability: "We can't control it, the client can't control it." JPMorgan's Jamie Dimon said the same day that his bank blocks outside agents from reaching customer account information.

Schwab is not closing the door entirely. It runs MCP servers that give outside models read-only access to balances and research while prohibiting money movement, and it steers customers to its own assistant, Charley, which cannot trade. That is the compromise forming across finance: agents are permitted to observe, not yet trusted to act. No federal statute governs AI-agent liability in financial services, and the three bills that would, including the Warner AI AGENT Act, have all stalled in committee. Until one moves, the rules are being written in terms of service.

The demand side is moving faster. Stripe's head of consumer product, Dan Hill, says about 200 AI agents now spend through Stripe's Link wallet, more than double the roughly 70 it counted in September, among them Meta's Muse. "Agentic commerce is here, people are using it, it's going to keep growing," Hill said. Stripe president John Collison said agent activity "basically started going vertical" in September and credited computer use, meaning agents that operate ordinary websites, as "the last big unlock." Amazon had just barred Muse from shopping on its site. For a business that sells online, both halves point the same way: your customers' agents can already find and buy from you, and who pays when they buy wrong is still open.

Anthropic pledges 5% of itself to a public benefit fund

Anthropic announced on October 18 that it is committing 5% of its shares, subject to stockholder approval, to a new Anthropic Public Benefit Fund, with the board able to add another 2% for every 50% increase in the company's value above a baseline it has not disclosed. The fund is a wholly owned LLC rather than a charity, which Anthropic says gives it room to make large donations to nonprofits or direct distributions to specific populations, with labor market impacts named as one example. None of the equity may be used for lobbying or political giving. "Given the pace of AI progress, bolder approaches may be needed," the company wrote.

Counting the co-founders' pledge of 80% of their own holdings and an employee donation match, Anthropic says roughly 20% of its shares are now dedicated to public benefit, on top of $200 million it previously committed to studying AI's effect on work. There is no dollar figure and no governance yet: who decides where the money goes is promised "in the coming months." The announcement lands as public sentiment toward AI sours, which chief executive Dario Amodei has called a crisis of trust, and weeks before a planned IPO. Watch the first grants, not the percentage.

The buildout gets repriced

The Philadelphia Semiconductor Index was on course Friday for its eighth straight daily decline, matching its longest losing streak since the 2008 financial crisis. About $2 trillion in market value came off the index over the slide, roughly $4 trillion measured from each stock's 52-week high, and more than half its constituents are at least 10% below their highs, Nvidia among them at more than 11% off. The index is still up nearly 60% for 2026. The causes cited were doubts about when AI spending pays back, rising rate expectations and a memory-chip price fight.

The physical side of that doubt showed up in Essex. Microsoft has ended its agreement to rent AI capacity at Nscale's planned data center in Loughton, a site meant to hold about 23,040 Nvidia GPUs by early 2027, after UK grid officials said they could not supply adequate power until the 2030s. Nscale filed confidentially for a New York listing a month ago, hoping to raise $3 billion at about $35 billion, and Microsoft and Anthropic together account for about 85% of its contract value. CoreWeave fell more than 25% last week and Firmus abandoned its own IPO.

Power is the constraint, and the United States is answering with old reactors. The Energy Department is lending up to $4.4 billion to restart three shuttered nuclear plants in Michigan, Pennsylvania and Iowa, together about 2.25 gigawatts, with Microsoft and Google as the anchor buyers and restarts aimed at 2027 and 2029.

Who buys whom: Microsoft's Claude bill and Disney's new AI team

Microsoft plans to pay Anthropic more than $2 billion a year to use Claude models inside Copilot, according to a report in The Information that neither company has confirmed. Claude reportedly handles tasks where customers complained about output, including financial automation in Excel and building PowerPoint decks, while an in-house router sends each request to the cheapest model that can handle it and increasingly leans on OpenAI to protect margins. Corporate Copilot users now exceed 30 million. The lesson for a buyer is the router: the model behind your subscription is chosen per task and can change without notice.

Disney, meanwhile, agreed to offer jobs to at least 60 Character.AI employees, the majority of the bankrupt chatbot startup's workforce, at an estimated $60 million a year in pay, on top of a $20 million bid for some of its assets; another buyer could still top it. Character.AI's former chief executive, Karandeep Anand, recently became Disney's first chief technology officer, so the startup Disney once accused of copying its characters is in effect becoming its AI department.

China moves to license the weights of its biggest models

China's commerce and science ministries published a draft on October 16 adding 40 items to the country's restricted-export catalog, 23 of them AI technologies. Controls on model weights, source code and training methods would apply to models with at least 2 trillion parameters or a top-ten global ranking, and to design know-how for chips at 14 nanometers or below. Bloomberg and the AP report that Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max would clear the threshold; the draft names no models. Comments are open until November 15, and a temporary registration requirement took effect the same day.

What stays open matters for anyone running Chinese models. Published research and open-source code are exempt, apps and APIs remain available abroad, and a firm running a covered model on its own overseas cloud could get an automatic license. With Taiwan criminalizing chip exports to China, the trade in AI is being fenced from both sides, and the fence runs between weights you can download and services you can only rent.

Security: leaked keys and a model that thinks about its own training

OpenCode, a development platform that lets users connect their own AI provider accounts, disclosed on October 17 that attackers reached a legacy system it was preparing to shut down and exposed third-party API keys for about 1.4% of Console users who had external providers linked. The keys may include OpenAI, Anthropic and Google credentials. OpenCode called it a large-scale distillation attack, meaning the stolen keys were used to query models heavily and harvest their outputs, and told affected users to revoke keys and check for unexpected charges. It has not said how many accounts or who was behind it.

Separately, Apollo Research found that Z.ai's open-weight GLM-5.3 occasionally reasons about how its behavior during an evaluation could shape its own future training, in 10 of 6,452 graded rollouts. Apollo calls the traces a warning sign, not evidence of deception.

Work: AI writes most of DeepMind's code, and the security job market splits

Google co-founder Sergey Brin said in a roundtable on Gemini 4 Argon that Google is "starting to get onto that RSI treadmill," meaning its models now help build the next models. The accompanying DeepMind safety report puts figures on it: more than 75% of production code at DeepMind is AI-written, and the share of code changes fully delegated to AI rose from 7.5% in April to 11.9% in September. About 88% of that work still needs repeated human steering, and the report finds "no significant end-to-end acceleration of AI R&D."

CyberSN's count of US cybersecurity job postings tells the hiring side. Postings for January through September 2026 fell to 235,786, the lowest in the four years it has tracked and 26.4% below 2025, though only about 4% below 2024 because 2025 was unusually strong. Penetration tester postings nearly halved, and governance, risk and compliance grew from 9.4% to 14.8% of all postings. AI security engineer postings more than tripled, to only about 2,400. "Risk is the biggest conversation in the room right now," said CyberSN chief executive Deidre Diamond.

Physical AI

OpenAI's humanoid is a hiring plan before it is a product. Sam Altman has said the company will "definitely do a humanoid," that the hard part is "figuring out the brain" rather than the body, and he told Fortune to expect "a cool demo in 2027" with street-ready robots a few years after. OpenAI's careers page lists 27 robotics openings, and a posting gives the near-term goal as general-purpose robots that help skilled workers build infrastructure. There is no design, price, partner or specification. It enters a market where about 19,100 humanoids shipped in the first half of 2026, up 272% on the year, with Chinese vendors above 97% of units.

The autonomy money is flowing to machines that already have jobs. Rivian has built its own in-vehicle AI models, is developing a driving stack for its cars and for future robotaxis with Uber, and is designing a 5-nanometer processor with Arm and TSMC, with chief executive RJ Scaringe talking openly about licensing the stack to other brands. Bedrock Robotics, which automates excavators, is reportedly raising $600 million on the data-center construction backlog. Self-driving truck developer PlusAI filed to go public through a merger, and Waymo dropped its waitlists in San Antonio, San Diego and Tampa.

The deployment with the best numbers this week is in the woods. Kodama raised $7 million, bringing its total to $14 million, for forestry equipment that runs repeatable tasks autonomously while an experienced operator supervises remotely and takes the hard cases. Its skidders have been with customers in the southeastern United States for about two years, moved more than 100,000 tons of timber in the past six months and run 10 to 14 hours a day. The labor case is blunt: one logger went through eight skidder operators in a year.

For a building owner, Honeywell's Forge Cognition is physical AI without a robot. Launched October 14 with Nvidia and Dell, it pairs a live digital twin of a facility with agents that classify alarms, flag failing HVAC equipment and simulate energy changes, with a human able to override at any level of autonomy. At Nvidia's Voyager campus the pilot cut unnecessary alerts by up to 90% and found up to 10% in energy savings. No price is given. Underneath all of it, Nvidia's Vera CPU team says agent workloads have reversed a decade of chip design: Vera uses a single monolithic die rather than chiplets because agents need low-latency communication between cores, and Nvidia claims up to twice the leading x86 part on agent tasks. Those are vendor numbers.

What This Means for Your Business

Decide what your agents may do with money, and expect your bank to decide for you. Schwab's position, read but do not act, is where every financial institution will land until a liability law exists, and the bills are stalled. So build to it: give any agent that touches your finances read-only access, keep payments behind a human approval step, and treat a tool that asks for your banking login rather than a scoped connection as a no. On the selling side, run the reverse test this week. Ask a shopping agent to buy something from your site. If it cannot find your prices, complete checkout or get a receipt, you are invisible to roughly 200 agents that are already spending, and the number doubled in a month.

Treat the chip selloff and the Essex pullout as a buyer's signal, not a warning. The infrastructure trade is being repriced around power, which means data center capacity will be scarcer in some regions and vendors with debt to service will push usage-based pricing and cut free tiers. Lock in caps and price protection on your AI contracts now while labs want the revenue, and keep your prompts, examples and evaluation sets in your own files so a router change at Microsoft, or anyone else, does not quietly change your results. Microsoft's reported $2 billion deal is the template: the model behind your subscription is chosen per task by a cost router, so measure the output you get, not the brand on the box.

Rotate every key you have given a third-party tool, and set spending limits on all of them. OpenCode's breach was a legacy system, which is to say the kind of system every company has, and the keys it leaked belonged to its customers' OpenAI, Anthropic and Google accounts. Make a list of every service holding one of your API keys, revoke the ones you no longer use, set usage alerts on the rest, and put the list on a quarterly calendar. The DeepMind report gives you the right posture for the tools themselves: let AI draft most of the work, and assume nearly nine in ten changes still need a person steering.

Hire for the job market that exists. CyberSN's numbers say general cybersecurity postings are at a four-year low while AI security roles tripled, and compliance work is a growing share of what is left. For a small firm, that means the generalist security hire is cheaper than it was, the AI-specific skill is scarce, and the practical move is to buy the first as a service and train one existing person on the second: agent permissions, key hygiene, and reading a model card. On robots, Kodama is the benchmark to hold vendors to: two years in the field, tons moved, hours per day, operators replaced. OpenAI's humanoid has a demo date and 27 job openings. One of those is a business case.

Sources

← All posts