Agents of Work
Let's Talk
October 18, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 18, 2026

Summary

Anthropic surveyed 10,980 Americans and its own staff and found the public expects AI to take about a quarter of their work by 2030, while the people building it expect three quarters. The same weekend Meta deleted the clause that let it crawl business websites, and advertisers learned they can target ChatGPT users by phone number. The distance between what AI builders expect and what customers will accept is now the planning problem for every business owner.

Highlights

  • Americans expect AI to do about a quarter of their work by 2030; the people building it expect three quarters. Plan for the middle.

  • Meta deleted the clause that let it crawl your website, but keep the robots.txt changes you made anyway.

  • Advertisers can target ChatGPT users by uploading your email or phone number, and the opt-out does not cover the current chat.

  • Two large open models landed in two days, one under a no-strings license; test one against your current AI bill.

  • GitHub Copilot removes Claude Opus 5 and two other models on November 16, so check any automated workflow now.

  • A credit agency rated Anthropic one notch below investment grade on limited disclosures, and the AI buildout runs on debt.

  • Workday's CEO says payroll and junior audit work goes first within five years, and hands-on "doers" will be fine.

  • Claude redesigned quantum circuits largely on its own, cutting one design's gate count 17-fold.

  • Palletizing robots now fit in a quarter less floor space, and Stryker put $400 million into a surgical robot maker.

  • Regulators are reviewing claims that Tesla robotaxis speed in school zones, with about 1,000 vehicles in scope.

Quick Takes

  • GitHub Copilot drops three models on November 16. Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra leave every Copilot surface, including agent mode and completions. GitHub suggests Opus 5.5 for the Claude models and GPT-6.1 Sol or GPT-6 Luna for Terra; annual individual subscribers keep Opus 4.8. Replacements switch on automatically for Business and Enterprise accounts unless an admin turned off the global default. Read more

  • An FDA-cleared AI reads the MRI you already have to predict breast cancer's return. SimBioSys's TumorSight Risk scores recurrence risk from the diagnostic MRI before surgery. In records from 2,129 women, 95.5% of those it called low risk were cancer-free five years later against 89.6% of those it called high risk. Oncologist Poornima Saha of Endeavor Health says it adds to established tests rather than replacing them. Read more

  • A model that looks again instead of guessing. Elorian, a Palo Alto startup led by former DeepMind researcher Andrew Dai, released Glimpse, which crops regions, pulls video frames and runs code mid-reasoning rather than converting an image to text once. The company raised $55 million at a $300 million valuation before shipping anything and prices the model at $0.60 per million input tokens and $3 per million output. Its benchmark claims are its own. Read more

What's Covered in Featured News

  • The AI expectations gap: what 10,980 Americans and 1,334 Anthropic employees expect by 2030, and what Workday's CEO says goes first.

  • Meta backs down: the crawling clause Meta removed from its business terms, what it says now, and what still applies.

  • ChatGPT ads by phone number: how advertisers match your contact details to ChatGPT accounts, and what the opt-out misses.

  • Two open models in two days: DeepSeek's new Flash with a search API and Reflection's 501-billion-parameter Beam.

  • The money and the chips: Anthropic's junk-grade credit rating and Taiwan's plan to criminalize AI chip exports to China.

  • Claude designs quantum circuits: eight tasks, one 53-hour run, and what the researchers had to correct.

  • Physical AI: Boston Dynamics' four-finger hand, palletizers for small factories, Stryker's surgical bet, robotaxis in school zones and the drone fight.

Featured News

Americans expect AI to take a quarter of their work. Its builders expect three quarters.

Anthropic published a comparison on October 16 of what ordinary Americans and its own employees expect from AI, and the two groups barely share a planet. Economists including Charles I. Jones and Anton Korinek surveyed 10,980 U.S. adults through Morning Consult between August 11 and 23 and polled 1,334 Anthropic staff the same week. Both groups agree on what AI can do today. On the future they split: 39% of adults expect AI to make a Nobel-level discovery by 2030, against 91% of staff. Among adults who use AI at work, 26% say a task now takes a quarter of the time or less; among staff, 59% do.

The number for a business owner is the one about work. Anthropic employees expect AI to handle about 75% of their current working hours by 2030. American workers put it at 28%, and only one in twenty expects 80% or more. Adults expect 42% of AI-capable tasks to be done by AI by 2030; staff expect 65%. Fed into Anthropic's economic model, the median adult's answers imply 5.4% GDP growth in 2030 and the median employee's imply 17.1%. Asked what blocks AI at work, Americans most often named customers preferring to deal with a person, then data privacy and security. A September follow-up with 7,613 adults produced the same picture.

Workday chief executive Aneel Bhusri gave Fortune the operator's version of the staff view on October 15. He expects "a lot" of displacement within five years, starting with payroll, junior auditors whose work Workday's audit agent now does, and the subset of lawyers its contract agent can replace; Workday is flattening its own management layers for the same reason. His safe list: "If you're a doer, an individual contributor, you're going to be fine," plus people who learn the tools and people who manage agents. He backs RAISE US, a retraining effort founded by former governors Gina Raimondo and Eric Holcomb, and says the companies profiting from AI owe the retraining.

Meta removes the clause that let it crawl your website

Meta has dropped the provision in its updated commercial terms that, as we reported Friday, would have let it crawl any website tied to a business account from October 30 and use the content to train its AI. The deleted language allowed Meta to "use any content from [a business account's] domains, web pages, or URLs" to provide, develop, improve, research and train its products. Legal experts interviewed by Nieman Lab had called it "a remarkable provision," and The Verge reported the removal; TechCrunch confirmed it October 17.

Meta's statement does not concede much. "The crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves," the company said, adding that the change "does not change how Meta uses web crawlers." A spokesperson said the same thing before the clause came out, which is the tell: Meta still crawls the open web for training, and a publishers' lawsuit filed this year will test what it may take. What changed is that your Facebook page no longer counts as permission. Friday's advice stands: keep the robots.txt entries for Meta's crawlers if you want them out, and keep anything you would not want in a training set behind a login.

Advertisers can find you in ChatGPT by phone number

OpenAI's ad system lets an advertiser upload a list of email addresses, phone numbers or Android advertising IDs, raw or hashed, and show ads to the ChatGPT accounts that match. 404 Media found the option by running two campaigns of its own, and Gizmodo confirmed the mechanics on October 16. A list must contain at least 25,000 entries before it can target, so one person cannot be singled out, and OpenAI says it does not tell the advertiser which accounts matched and deletes the uploaded file within 24 hours. It did not answer whether it hashes raw data on arrival.

The opt-out is narrower than it looks. Turning off personalized ads stops OpenAI using "your past chats, activity, and preferences to select ads," but its own wording continues: "Ads may still be based on your current chat." For a small business that advertises, this is a customer-match tool like Meta's and Google's, with the same obligation to have consent for the list you upload. For a small business whose staff use ChatGPT, whatever they type into the free tier is ad inventory.

Two open models in two days

DeepSeek released V4.1-Flash-1017 on October 17 with weights on Hugging Face and ModelScope, calling it a "Pareto improvement in performance and cost" over the September Flash. The model now sits behind the deepseek-flash API name, and retired model names route to it automatically. The bigger news for builders is a Search and Fetch API backed by DeepSeek's own search engine at $0.75 per 1,000 calls per endpoint, free through Claude Code, Codex and DeepSeek's own harness until October 31. DeepSeek published benchmark charts, not numbers.

Reflection AI, the Nvidia-backed startup, made good on its promise the day before: Beam, a 501-billion-parameter mixture-of-experts model with 23 billion active, is on Hugging Face under Apache 2.0 with no usage restrictions, in three precision formats and with day-one support in vLLM, SGLang and Ollama. It is available through AWS, Google Cloud, Microsoft Foundry and the usual inference providers, Dell and HPE sell it for on-premises and air-gapped use, and Reflection's own platform is free in beta. The announcement gives no benchmark scores; they live in the model card, and the company says Beam performs best inside its own Mirror harness. With Meta's Muse Spark Open on Wednesday, that is three large open models in a week.

The money and the chips

Anthropic has carried a BB+ credit rating from Kroll Bond Rating Agency since June, one notch below investment grade, a fact that stayed private until the Financial Times reported it this week. The rating was issued for a $35 billion financing led by Apollo, with BlackRock participating, that funds Broadcom's chip platform for Anthropic. About $4.4 billion of that is subordinated debt exposed directly to Anthropic's credit at a yield near 8.5%, against 5.75% on the senior debt. KBRA worked from limited disclosures: Anthropic shared some revenue figures but not free cash flow or EBITDA, so the agency used its own scoring system rather than a full model. Bankers are reportedly lobbying for an investment-grade rating before the planned IPO. Alongside the $200 billion in private AI loans we covered yesterday, the buildout's financing is the story to watch.

Taiwan, meanwhile, plans to make unauthorized exports of advanced AI chips and servers to China a criminal offense. Economics minister Kung Ming-hsin told Bloomberg the items will be added to the controlled list under the Foreign Trade Act, partly to meet commitments in Taiwan's trade agreement with the United States. In August, prosecutors indicted nine people over 130 AI servers, 74 of them allegedly routed to China through other countries, but could only charge breach of trust and document forgery because AI hardware was not specifically covered. No effective date has been set.

Claude redesigns quantum circuits, mostly on its own

Anthropic's science team reported October 16 that Claude Mythos 5, given plain-language specifications and an automated checker that scored designs by qubits times gates, produced circuits for eight computational tasks that beat the best published versions. Andi Gu led the work in Anthropic's STEM Fellows program with Alek Kemeny, and Harvard's Susanne Yelin and Mikhail Lukin collaborated. In a 53-hour run on modular inversion, Claude found a reversible version of a known method that avoids storing branch history; one resulting circuit uses 12% more qubits than the smallest prior design but 17 times fewer gates. On a decoder central to a Google optimization algorithm it cut the gate count 12-fold and reserved 19 qubits where a naive design needed up to 840. The caveats are instructive: the model was initially too cautious, re-checking settled results and favoring familiar methods, until the researchers forced it to pursue several approaches in parallel. Circuit design is one piece of a working quantum computer, and the paper claims nothing about timelines.

Physical AI

The most useful robotics decision of the month was to remove a finger. Boston Dynamics' new hand for Atlas has four fingers and 13 degrees of freedom, up from seven, and it dropped the pinky after chief product and technology officer Zachary Jackowski had the team tape their pinky and ring fingers together for a day. Engineer Dylan Thrush said the extra dexterity was not worth "the extra complexity" of three more joints, plus size, power and actuator cost. The hand is sized like a large human hand so human demonstration data transfers, and is "designed for mass manufacturing," though no price, weight or production volume is given. At the Hyundai plant where Atlas is being trained, more than 1,000 robots already work beside nearly 1,500 people; Atlas sorts roof racks, with permanent deployment "several years" away.

For a small manufacturer or distributor, the week's news was end-of-line equipment that fits. At PACK EXPO in Chicago, Scott Technology launched the NexPAL 200HL, a palletizer about 25% smaller than comparable machines that handles more than 35 cases a minute, with an augmented-reality view that drops a full-scale model into your plant from a QR code; its first order was four systems for a Georgia poultry processor. CMES Robotics made the U.S. debut of a mixed-case palletizer whose vision system stacks cases of varying sizes on one pallet, aimed at 3PLs and omnichannel warehouses. NAPA Auto Parts chose a Hai Robotics storage system through integrator Hy-Tek after testing it in Hy-Tek's lab. None of the three disclosed a price.

In surgery, Stryker invested about $400 million in CMR Surgical, took an exclusive option to buy the rest, and became exclusive U.S. sales agent for the Versius robot, which is modular, needs no dedicated operating room, and has been used in 45,000 cases in more than 30 countries. Versius Plus has FDA clearance for gallbladder removal, with gynecology pending. For a surgical center, the pitch is a robot that fits an existing room and a Stryker rep who already visits.

On the road and in the air, the regulators are busy. NHTSA opened a review file on a petition alleging Tesla's robotaxis speed through active school zones, citing four incidents from social media video, no crashes, and an estimated 1,000 vehicles in scope, the first official hint of fleet size; it has until about January 21 to decide whether to investigate. The FCC placed classified Pentagon letters into its DJI and Autel proceedings, though it says it did not rely on them and the Covered List still does not ground drones people already own. And at a Beijing counterterrorism forum with delegates from 62 countries, China showed low-cost AI-guided interceptor drones built to knock down small drones, with no specifications released.

What This Means for Your Business

Plan for the middle of the expectations gap. Anthropic's staff may be right that AI does three quarters of knowledge work by 2030, and your customers may be right that they still want a person. The survey's own finding about barriers is the operating rule: customers prefer dealing with a human, so automate the work behind the counter first and keep a person at the point of contact. Bhusri's list is a decent map of where to start: payroll, routine audit checks, first-pass contract review. His other point matters more for a small firm: the people who do the hands-on work are safe, and the people who learn to manage agents get more valuable. Decide which of your staff become agent managers and give them the time to do it.

Take Meta's reversal as a reprieve, not a resolution. The clause is gone, but Meta's stated position is that nothing about its crawling changed, and the publishers' lawsuit will decide what it may take from the open web. Leave your robots.txt entries in place, keep client portals and pricing behind a login, and read the next terms update when it arrives rather than when a lawyer flags it. Apply the same reading to ChatGPT's ad system: if you advertise, you now have a customer-match channel that needs the same consent you would need for Meta or Google; if your staff use the free tier, assume the conversation is ad inventory and move anything sensitive to a paid or self-hosted option.

Use the model churn to cut your bill. In one week you got Meta's Muse Spark Open, Reflection's Beam under a license with no revenue conditions, and a cheaper DeepSeek Flash with a search API at 75 cents per thousand calls. Pick your two most expensive AI jobs, run them on one open model through a hosted provider, and compare on your own documents, not vendor charts. Do the same check in reverse for anything pinned to a model that is going away: GitHub Copilot drops Opus 5 and Terra on November 16, and any prompt-tuned workflow in your CI pipeline will fail silently if nobody moves it. Keep prompts, examples and evaluation sets in your own files so the next switch costs an afternoon.

Watch the financing, not the headlines. A junk-grade rating on limited disclosures, debt at 8.5%, and a chip-export rule that can criminalize a shipment are the mechanics underneath the AI prices you pay. None of it changes your quote this month, but a lab that must satisfy bondholders will raise prices or cut free tiers before it misses a payment, so negotiate caps now while buyers have leverage. On the robot side, the right question for a palletizer, a storage system or a surgical robot is the one none of the vendors answered this week: the price, the uptime and the intervention rate. Ask for a reference customer's numbers, and if the answer is a demo, wait.

Sources

← All posts