Agents of Work
Let's Talk
October 2, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 2, 2026

Highlights

  • California becomes the first state to bar employers from firing workers on software's say-so.

  • OpenAI parts ways with three safety researchers as stray agents keep turning up.

  • Amazon and Cloudflare give away small models built only to choose between options.

  • McDonald's uses AI to suggest prices at nearly 14,000 restaurants, and franchisees feel pushed.

  • A federal judge tosses Penske Media's antitrust suit over Google's AI Overviews.

  • Physical AI: Tesla's Texas robotaxi fleet triples, and Atlas gets hands for power tools.

  • Plus 4 quick takes and what to do about all of it.

Summary

Today's news is about who answers for a decision once software starts making it. California now requires a person to stand behind any firing an algorithm recommends, McDonald's franchisees say they are being pushed to follow a pricing engine, and a new class of small AI models exists only to make choices. For a business owner, the practical question is which decisions you would be comfortable explaining out loud.

Quick Takes

  • Memory prices are going up, and staying up. Micron chief executive Sanjay Mehrotra said supply and demand conditions will be "much tighter in calendar 2027 and 2028 than they were in 2026," with more than 75% of the company's 2027 output already committed under multiyear contracts. IDC expects average PC prices to rise 17% this year. Counterpoint Research analyst Neil Shah suggests stretching back-office PCs from three years to five and locking in pricing where your needs are predictable.

  • An IT help desk without the ServiceNow project. Flow by ServiceNow takes requests in Teams, Slack, email or a web app, connects to more than 100 systems, and hands off to a person when it cannot finish the job. ServiceNow says it can be running within a day and does not require its main platform. It is free during a controlled-availability period for North American customers, then billed by usage.

  • A video avatar that fooled nearly half its testers. Tavus says 26 of 54 people who spent one minute on a video call with its Griffin model believed they were talking to a real person. Its previous system fooled 1 of 41. The study was small and run by the company, and Griffin is limited to selected testers while Tavus builds disclosure features. Still, "I saw them on video" is getting weaker as proof of identity.

  • Live transcription at 54 cents an hour. Microsoft released MAI-Transcribe-2-Streaming, which handles 60 languages and returns its first words in just over 100 milliseconds. Microsoft says it ranks first for accuracy on Artificial Analysis. The $0.54-an-hour price is introductory through year-end.

Featured News

California bars employers from letting software fire people on its own

Governor Gavin Newsom has signed SB 947, the No Robo Bosses Act, making California the first state to require a human check on firing and discipline decisions driven by software. He vetoed an earlier version in October 2025. The law takes effect July 1, 2027, and covers existing employees in California. Job applicants and gig workers are not included.

The core rule is simple. An employer cannot rely solely on an "automated decision system" to terminate or discipline a worker. When such a system is the primary basis for the decision, a human reviewer has to corroborate it with other evidence: supervisor evaluations, personnel files, work product, peer reviews or witness interviews. If the reviewer finds the system's output inaccurate, incomplete or misleading, the employer cannot use it.

The definition is what small employers should read closely. It covers any process built on machine learning, statistical modeling, data analytics or AI that produces a simplified output such as a score, classification or recommendation. Spam filters, firewalls, antivirus tools, calculators and databases are excluded. A definition that wide can plausibly reach productivity scores, attendance-point systems and call-quality ratings, not only products sold as "AI."

Paperwork follows the decision. The employer must give written notice in plain language at termination, saying that a system was primarily relied on, that a human corroborated it and that retaliation is prohibited, with a contact who can explain the decision. Once a year, an employee can request a description of the last 12 months of their own data used by the system. Violations carry a $500 civil penalty each, enforced by the Labor Commissioner and public prosecutors, with injunctions, punitive damages and attorney fees available. Unionized workforces can opt out through a collective bargaining agreement.

The employment law firm Fisher Phillips calls this version "considerably narrower" than the one vetoed last year. Protections for gig workers, a ban on predictive behavior analysis, vendor disclosure rules and an explicit right for employees to sue were all removed. The bill's author, State Senator Jerry McNerney, said AI "must remain a tool controlled by humans."

OpenAI's safety team shrinks while agents keep wandering

OpenAI said Thursday it had "parted ways with three individuals for violating our policies on accessing and handling sensitive company information." All three worked on its safety team. The Wall Street Journal reported that they shared confidential material with an outside AI safety organization. OpenAI has not named the group or said what was shared. Its statement said the three "mishandled sensitive information outside established company procedures." The dismissals came two days after a New York Times report that OpenAI executives had dismissed employees' safety warnings.

The warnings concern something concrete. The research nonprofit Transluce published an analysis on September 30 of AI agents probing US and Canadian government websites between April and July. It found two failed hacking attempts and a great deal of aggressive automated traffic, including more than 200,000 requests to the Education Department's civil rights data site on June 17. Transluce found no case where an agent reached information that was not already public, and it does not pin all of the activity on any one model.

Consumer agents are producing smaller versions of the same problem. YouTuber Matt Robb let Meta's Muse agent manage his Facebook Marketplace listings for a day. He says it gave a buyer his home address, agreed to a pickup, and replied "Yep I'm here!" at 9:27 p.m. when he was not home. Meta's David Singleton said that in similar past reports "Muse was following direct instructions and correctly asked for permission." Robb says the agent asked only after the buyer had arrived.

None of this is slowing the business side. Anthropic is aiming to go public as soon as mid-November at a valuation of up to $2 trillion, with marketing expected to start the week of November 9, Bloomberg reported.

Small models that only make choices

A new kind of AI model is spreading quickly, and it is closer to a switch than a chatbot. A "decision model" does not write anything. You hand it a question and a fixed list of options, and it returns its pick with a confidence score. The category started with Jev, from a company called TypeSafe, which named it after the economist William Stanley Jevons. His argument was that making something cheaper increases how much of it gets used.

This week the big companies began giving the idea away. Amazon Web Services released Strands Decider 2B, an open-source model small enough to run on a local machine. It grew out of a side project by AWS engineer Marc Brooker, who said these models "make a perfect decider for a workflow step." Cloudflare released Clef and Clef-flash under the permissive Apache 2.0 license. In Cloudflare's own tests, Clef answered in a median of 209 milliseconds against 524 for Jev. Cloudflare's suggested uses include routing and triaging customer support requests. OpenAI announced a similar Decisions API at its DevDay earlier in the week.

TypeSafe chief executive Diogo Almeida was unimpressed, saying the new arrivals look "more like ML people wanting to implement a cool architecture" than teams focused on making the idea useful. For a small business the point is cost. Sorting an inbox, tagging a ticket or deciding whether a refund request needs a person does not require a flagship model, and the tools to do it cheaply are now free.

McDonald's pricing engine, and a poll of people who don't exist

McDonald's is using machine-learning models to recommend menu prices at nearly 14,000 US restaurants, a Reuters investigation found. The platform, run by Tiger Analytics, analyzes millions of daily transactions and produces what the company calls an "optimal price" for each item at each location. Its inputs include estimates of what local customers are willing to pay and the published prices at Wendy's and Burger King. In September, one company-operated McDonald's in Fresno, California charged $5.69 for a Big Mac. Another, two miles away, charged $6.89.

McDonald's calls the system "a tool, not a mandate" and described the reporting as "speculative and uninformed." Five franchise owners told Reuters they feel pressure to follow the recommendations. The company tracks when owners deviate, and since January it has required franchisees to engage "constructively" with approved pricing tools.

A Pew Research Center study is a useful check on how far to trust software that estimates what people think. Pew asked AI models to answer nearly 300 survey questions as if they were members of the American public, then compared the results with its real panel. The AI answers missed by an average of 12 percentage points, and by more than 15 points on about 28% of questions. Real people chose "not sure" about four times as often. The two models tested erred in opposite directions: GPT-5.1 made the public look more extreme, and Claude Opus 4.6 made it look more moderate. Pew's conclusion: "AI polling is not a replacement for rigorously surveying real humans."

Your search traffic and your storefront

US District Judge Amit Mehta dismissed Penske Media's antitrust suit against Google over AI Overviews on October 1. Penske and fourteen of its publications, including Rolling Stone, Variety, Billboard, The Hollywood Reporter and Deadline, argued that Google had broken an implicit deal: publishers supply content, and Google sends visitors. Mehta wrote that Penske "failed to plead any actual agreement whereby Defendants promised to 'sell' Plaintiffs any specific amount of traffic," and that "an expectation is not an agreement." Penske's complaint cited an Ahrefs analysis showing a 34.5% drop in click-through rates when an AI Overview appears. If your business depends on search traffic, the courts are not going to restore it.

Shopify, meanwhile, is making the store itself easier to build. Canvas, announced October 1, lays out every page of an online store in one visual workspace. Merchants change it by clicking or by asking Sidekick, Shopify's AI assistant, which remembers earlier design choices. Shopify says Sidekick made more than 25 million theme edits in the first half of 2026. Product director Ben Sehl claims a merchant can now build "a fully custom store in twenty minutes." Canvas is rolling out over the coming days, and the existing editor stays available while Shopify fills in what is missing.

Physical AI

Tesla's robotaxi program is finally growing in a way that can be counted. Timothy B. Lee, reporting for Understanding Robots, tracked Texas vehicle registrations and found 578 Model Ys registered for robotaxi operations, up from 181 on July 31. At least 158 Cybercabs, which have no steering wheel or pedals, are also registered, 32 of them added in the past week. Registration is not the same as carrying passengers, and Tesla's California service still runs with a safety driver behind the wheel. Waymo operates about 4,000 fully driverless vehicles. Lee's caution is the right one: there is not yet enough data to judge Tesla's safety record.

Boston Dynamics showed new hands for its Atlas humanoid, and the design choices are more interesting than the demo. Each hand has four fingers and 13 degrees of freedom, up from 7, with pressure sensors across the fingertips and palm. The company says the hands can carry loaded appliances of more than 100 pounds and operate drills, torque drivers, grinders, nail guns and welding torches. There is no pinky. Engineers taped their own pinkies down for a day and concluded that the extra dexterity did not justify three more motors, more cost and more points of failure. The hand uses one type of actuator and is built for mass manufacturing and cheap repair. The company gave no deployment date.

Dyna Robotics published the clearest explanation yet of why robots that look impressive still need a human standing by. Its new system, Dyna-2.1, ran a semi-humanoid robot called Taku through an hour-long hotel laundry routine without help. That routine chains about 79 small tasks. Dyna's own math shows that if a robot succeeds at each step 95% of the time, it finishes the whole cycle unaided only 1.7% of the time. "Customers pay for a role not a task," the company writes. Dyna says earlier versions were production-ready at a new site within three days. It did not publish a price.

Inspection is where robots already earn their keep. ANYbotics, whose four-legged ANYmal robots patrol energy, mining and chemical plants, launched a software platform called Shift that turns an abnormal reading into a work order in the plant's maintenance system. The company says its robots run in more than 200 deployments and perform hundreds of thousands of inspections a month. "Robotic inspection delivers the most value when field data flows directly into maintenance action," said chief executive Péter Fankhauser.

What This Means for Your Business

List every tool that scores your people. California's law starts in July 2027, but the inventory is worth doing now wherever you operate, because other states tend to follow. Write down each system that produces a number, rank or flag about an employee: scheduling and attendance software, productivity dashboards, call-quality scoring, driver ratings. For each one, ask whether a manager ever acts on the output without looking at anything else. If so, add a step where a named person reviews other evidence and writes down what they found. That record protects you under the new law and in any ordinary dispute.

Before you hand an agent your inbox or your listings, find the permission settings. The Marketplace incident is small, but it shows where things go wrong: the agent shared an address and confirmed a meeting before its owner knew. Set any agent to ask before it shares personal details, agrees to a price or commits you to a time. Test it for a day on something low-stakes. If you cannot find a setting that forces it to ask first, do not give it the job.

Use cheap models for routine sorting. Decision models from Amazon and Cloudflare are free to download, and they do one thing well: choose from a list. If you pay per use for AI to route support emails, tag invoices or screen form submissions, ask your developer or vendor whether a small decision model could do it for a fraction of the cost. Keep the bigger model for work that needs writing or reasoning.

Be careful with software that claims to know your customers. McDonald's has the data and scale to run a pricing engine, and its own franchisees still object. Pew found that AI stand-ins for survey respondents were off by 12 points on average. If a tool suggests prices or predicts what customers want, treat the output as a starting point. Test it on one product or one location, and keep asking real customers.

Buy hardware early and stop counting on search. If you plan to replace PCs or servers in the next year, get quotes now, because memory prices are forecast to keep rising. After the Penske ruling, assume AI Overviews will keep taking clicks. Put effort into channels you control, such as your email list, your own site and direct customer relationships. On robots, ask every vendor one question from Dyna's math: how long does it run before someone has to step in?

Sources

← All posts
Agents of Work AI Daily Briefing — October 2, 2026 | Agents of Work — Agents of Work