Agents of Work
Let's Talk
October 29, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 29, 2026

Summary

The AI tools businesses use are becoming more capable and more persuasive, and that makes the person checking the work more important. Google's new model is built to find and fix security holes, and an Anthropic study shows experts quietly adopting an AI agent's preferences. Vendors are also adding meters, budgets and kill switches to agents, and Delaware is weighing companies run entirely by one.

Highlights

  • Experts followed an AI agent's hidden lean 86% of the time, and few credited the agent for it.

  • Before your team asks AI for a recommendation, have them write down their own view first.

  • Google's Gemini 4 Carbon finds and patches software flaws, but small businesses cannot buy it yet.

  • Anthropic now shows enterprise customers what their AI spending produced; ask every AI vendor for the same report.

  • GitHub Enterprise customers can now pay for Copilot by usage instead of buying a separate subscription.

  • Salesforce's automatic shutdown for failing agents arrives in January; budget caps and monitoring are already common.

  • Delaware may let an AI agent run a whole company, and Congress is already asking who carries the liability.

  • North American robot orders rose about 25% last quarter, led by auto parts and electronics makers.

  • A signed humanoid order is not a working robot; ask robot vendors for hours of useful work.

Quick Takes

  • Microsoft's cloud keeps growing fast. Microsoft reported revenue of $92.3 billion for the quarter ended September 30, up 19%, with Microsoft Cloud revenue of $63.9 billion, up 30%, and Azure up 48%. Net income rose 32% to $36.6 billion. For buyers, it confirms that cloud and AI services are a fast-growing line in customers' bills, so budget for yours to grow too. Read more

  • Chrome's autofill can now pull details from Gmail. Once a user turns it on in Chrome settings, Chrome can suggest flight confirmation numbers, membership IDs, tracking and order numbers, addresses and even credit card numbers found in their email. It is rolling out to U.S. users on desktop and in the Gemini in Chrome apps for iOS and Android. Google says sensitive data is encrypted; TechCrunch's report does not say whether work Google Workspace accounts are covered, so check before staff switch it on for company inboxes. Read more

  • Google opens Gemini and Search to outside apps. Google's Connected Apps Platform lets businesses connect their services so users can take actions inside Gemini and Search, as they already can with Canva or Instacart. Gemini has more than 50 connected apps, including Adobe and Peloton, and Calm is joining both. The program is in beta for U.S. developers, with a waitlist for the public preview. Read more

What's Covered in Featured News

  • Google's Gemini 4 Carbon: a model built for finding, exploiting and patching software flaws, who gets it first and what it will cost.

  • AI agents steer the experts using them: what Anthropic found when analysts' agents were secretly set to favor one answer.

  • Metering and controlling AI agents: Anthropic's spending reports, GitHub's pay-per-use Copilot and Salesforce's agent kill switch.

  • Delaware's AI-run companies: what the proposed Artificial Intelligence Company would allow, and why Congress is asking questions.

  • Physical AI: North American robot orders, Maximo's solar construction robots, Agility's farm-machinery pilot and how to measure whether humanoids do real work.

Featured News

Google's Gemini 4 Carbon is built to find and fix software flaws

Google released Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon, and called it its strongest model. In the launch post, Koray Kavukcuoglu, Google DeepMind's SVP and Google's chief AI architect, describes it as built for long, multi-step engineering work and, above all, for cybersecurity. Google reports it scores 88.2% on SEC-bench Pro, a test of finding software vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching them. It also ranks first on Artificial Analysis's Intelligence Index v5.0 with a score of 59. Chrome's own security team told Google that Carbon was more than 10 times better than another lab's large model at writing high-quality patches, and in Google's internal tests it produced secure code at more than twice Argon's success rate.

Google is not releasing it to everyone. This week a subset of cyber defenders get access, with more trusted partners added gradually through what Google calls the Fairwind Program, and the company is taking part in the U.S. government's voluntary testing of models before release. Developers, businesses and consumers will get it "as soon as possible," with no date. When it does open up, introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 afterwards. Google also reports safety gains over Argon, including an almost 80% drop in high-severity misalignment alerts in its internal coding environments and zero successful attacks on Gray Swan's test of hidden instructions planted in content the model reads. All of these figures are Google's own.

The significance for a small business is indirect but real. It follows Anthropic's report this week that its research model chained bugs in Chrome and Safari into working attacks. The best AI models are now being judged largely on how well they find and fix security holes, and both labs are handing them to defenders first. Software will get patched faster, but only for people who install the patches. Automatic updates on every browser, phone and laptop are the one control any business can switch on today.

AI agents steer the experts who use them

Anthropic published a study on October 27 that every manager deploying AI assistants should read. Researchers recruited 61 experts in causal inference, a branch of statistics, and asked them to advise a fictional farming cooperative on whether to expand or end an advisory program. Fifty worked with Claude Opus 4.8 running in Claude Code. Each analyst made four analytical choices, each with two defensible options, and before the task a coin flip secretly set which option their agent would favor. The agents were told to argue honestly and change course if the evidence demanded it.

The analysts matched the agent's hidden lean 86% of the time, against about 50% if the agent had no influence, ranging from 81% to 92% depending on the decision. When the agent leaned toward rolling the program out, 54% recommended rollout, compared with 36% when it leaned the other way. Where analysts had formed a view before the agent spoke, they switched 60% of the time when it disagreed. Experience and education did not predict who deferred. Just over half said the recommendation was more their own than the agent's, and 77% repeated the agent's arguments in their write-ups, sometimes word for word, but only 15% credited it.

The agent also helped. Six of the 11 analysts who chose to work without it made serious methodological errors, against three of the 50 who used it, though the groups differed in experience. Anthropic notes the agents' leans were deliberately strengthened and that the study covered one task, one field and one model. The lesson for operators: an AI that drafts the analysis also shapes the decision, so the human check only works if people form their own view before reading the AI's.

Vendors add meters, budgets and kill switches for AI agents

Three announcements this week show AI spending moving from flat subscriptions to metered use, with controls to match. Anthropic launched reporting for enterprise customers that shows managers, for a chosen date range, what a team produced with Claude, what it cost and where its work got stuck, along with recommendations and pooled team budgets. Anthropic says it will not train on that data by default. The pressure behind it shows in figures from Jellyfish, an engineering analytics firm: spending on tools built on Anthropic's models has grown tenfold since January and now equals about 2% of total engineering cost. "That 10x growth is growth that most CFOs didn't plan for," said Jellyfish chief scientist Nicholas Arcolano.

GitHub, at its Universe conference, said Enterprise customers can now use all Copilot features without a separate subscription, paying only for the AI credits they consume; existing Copilot Business and Enterprise customers can keep their plans. Its new Copilot Fusion routes each coding task to one or more models, and GitHub says its top setting scored within two points of Anthropic's strongest coding setup at half the cost and 40% less time per task.

Salesforce's MuleSoft is extending Agent Fabric, which manages agents built on any platform, with dollar or token budgets by provider and model, and testing tools that enter beta in November. A kill switch that shuts down an agent automatically when it fails those tests is due in January 2027. "We're not moving the agent; the agent's still running there," said Andrew Comstock, MuleSoft's general manager, of how the registry discovers agents across platforms. Pay-as-you-go pricing makes AI easier to start using, but harder to forecast.

Delaware weighs companies run by an AI agent

Delaware, where most large U.S. companies are incorporated, is drafting a new kind of business entity: an Artificial Intelligence Company, run day to day by an AI agent rather than a person. Under the draft, an AIC "may have any lawful purpose" except banking, and the agent must log its activity. A single member, a person or a company, must keep it adequately capitalized, and is not liable for its debts unless it fails to do so or uses the company to break the law or commit fraud. The draft would also set up a 30-month regulatory sandbox. Secretary of State Charuni Patibanda-Sanchez and John Nay, CEO of legal AI firm Norm Ai, announced the plan in July, and lawmakers are expected to take it up in January.

On October 26, House Oversight Chairman James Comer wrote to Governor Matt Meyer asking for the state's risk assessments, warning the plan "could be used to protect powerful AI companies and their investors from responsibility." Critics' central question is who pays when an agent-run company breaks a contract or hurts someone. "We're trying to define where exactly liability lands," Patibanda-Sanchez has said. For businesses, the practical issue arrives sooner than the law: you may soon be signing contracts with a supplier that has no human manager.

Physical AI

Industrial robot demand is recovering, and broadly. The Association for Advancing Automation says North American companies ordered 10,989 robots worth $723.5 million in the third quarter, up 24.8% in units from a year earlier. Orders from automotive component makers rose 90.4% and from semiconductor and electronics makers 97.6%, while food and consumer goods fell 23.7%. Collaborative robots, the smaller arms designed to work beside people and the kind most small manufacturers buy first, rose about 56% to 1,831 units, 16.7% of the quarter's orders. A3's Alex Shikany pointed to "how broadly it was distributed across the market." If you have been waiting for robot integrators to have spare capacity, demand is moving the other way.

Maximo Robotics offers a model worth studying. It says its robots have completed more than 300,000 solar construction tasks under performance-based robotics-as-a-service contracts, at 2.5 times the manual installation rate per worker per shift hour, on sites in seven states including Texas and Kansas. Customers pay for completed work rather than buying robots, which moves reliability risk onto the vendor. Primoris, one of its contractor partners, is integrating the system into field crews.

Humanoids remain mostly at the pilot stage. Agility Robotics and farm equipment maker AGCO signed an agreement for a Digit pilot at AGCO's Hesston, Kansas plant, beginning in 2027, sorting and stacking parts and placing totes. AGCO joins Toyota, GXO and Schaeffler on Agility's customer list. A recent Robotics 24/7 analysis by Lars Talbert shows how big the gap is between orders and work: Agility cites more than $300 million in multi-year Digit orders tied to milestones, yet its merger filing reports about $1.8 million in 2025 net sales. "A signed order, a delivered robot, and a robot doing useful work are separate milestones," Talbert writes, noting that no company publishes uptime or human-intervention figures.

China's AgiBot is proposing a fix. At its Innovation Day it introduced Worken, a measure that multiplies active robots by hours worked without human help by output compared with a human benchmark. "How much productive work are they actually delivering?" asked co-founder Peng Zhihui. It is a vendor's own yardstick, but the questions it asks are the right ones for any buyer.

What This Means for Your Business

Protect your team's judgment. Anthropic's study used experts, and they still followed the agent's hidden lean most of the time. For pricing, hiring, credit and vendor decisions, have the person responsible write down their recommendation and reasons before they open the AI's analysis. Better still, ask the AI to argue both sides, or to ask for the person's view first. Record which decisions were AI-assisted, so you can review them later.

Get a usage report from every AI vendor. As billing moves to credits and metered use, find out what each tool costs per team per month and what it produced. If a vendor cannot show you that, set a monthly cap with them and check the invoice against actual use. Start with coding tools and customer-service agents, which are where usage grows fastest.

Treat every agent like a contractor with a spending limit. Give each one a named owner, a budget, a way to switch it off and a record of what it did. Salesforce's automated kill switch is not due until January, but you can set the rule now: if an agent fails a check or exceeds its budget, a person decides whether it continues.

Keep patching. Google and Anthropic are pointing their strongest models at security flaws, which will speed up both patches and attacks. Turn on automatic updates for browsers, phones and operating systems, and think twice before letting Chrome's new autofill read a shared company inbox.

If you are buying robots, pay for work rather than hardware. Ask for performance-based terms like Maximo's, or for measured hours of autonomous work, not unit counts and order books. With robot orders rising, book integrator time early.

Sources

← All posts