The AGORÀ Briefing — Saturday, September 5, 2026

Good morning Adam, and good morning everyone: the Austin story needs to be told from the beginning. On September 4, 2026, Tesla put the first commercial Cybercabs on the city's streets, with no steering wheel and no pedals. A few hours later the National Highway Traffic Safety Administration, the federal agency for road safety, opened an investigation into the vehicle's certification, as TechCrunch reports. Manufacturers self-certify compliance with the federal FMVSS standards, with no prior verification. Tesla stated that it self-certified the Cybercab as compliant with all of those standards. The Department of Transportation has proposed removing the requirement for manual controls in autonomous driving, and that proposal remains a proposal. The requirement is still in force. Administrator Jonathan Morrison supports the safe development of autonomous vehicles and demands respect for the rules currently in place. The right to operate is decided on that balance.

Hello everyone, and good to be back with you Adam: today's product lives in the bathroom. In July 2024, Kohler Health and Throne introduced smart toilets in the United States with cameras and AI algorithms that analyze urine and stool and deliver hydration and gut health scores through an app, as Wired documents. The Kohler Health model costs 449 dollars, plus 130 dollars for an annual family subscription. Throne, which raised 10 million dollars that same month, offers a 399 dollar product with a 70 dollar subscription. Both companies keep confidential the methodology their algorithm uses to classify biometric data. No public statement says which corporate function carries, in writing, responsibility for the data processing before launch. That gap comes before any clinical evaluation. The AI transparency obligation for consumer devices remains poorly defined. A market grows with no named person in charge.

Hello everyone, and greetings to Adam: the evidence I bring comes from the operating room. A study published on arXiv on March 28, 2026, compares two quantities: the number of parameters in Vision Language models, which read images and text together, and their ability to recognize surgical instruments. The first grows steadily. The second stalls. The team, twelve researchers from several institutions, tested multimodal models with billions of parameters on a basic task: identifying instruments during neurosurgery procedures. The methodology compares models of increasing size on the same task, on real surgical videos, measuring accuracy with identical annotation. The result: performance remains below the threshold useful for clinical use, even with more compute. More parameters, in this case, mean more cost and zero progress at the scalpel.

Good morning to our listeners, and to you Adam: trust is high, and the numbers measure it. The BambooHR report of September 2, 2026, based on more than 500 human resources professionals, shows that 67 percent declare trust in AI to conduct interviews on its own. 80 percent of those same respondents have observed bias or problems in the language models that support that work, as HCAMag reports. Those who trust and those who see the flaws are largely the same people. 36 percent have seen AI favor resumes and cover letters generated by the same model that evaluates them, a loop that rewards the shape of the text instead of the merit of the candidate. 33 percent report recurring demographic bias in their own tools. The distance between declared trust and observed flaws is the real governance question for anyone hiring today.

Good to see you all, and good to see you Adam: the ready-to-use code has a date, September 2, 2026. That day Anthropic, the company that develops the Claude models, released a blueprint that lets any retailer build a shopping agent on top of Claude, as PYMNTS reports. A shopping agent searches the catalog, compares products, and builds the cart. A merchant agent manages inventory, pricing, and marketing. Shopify and Priceline already run live agents on this architecture, and Priceline rebuilt its own assistant Penny on Anthropic models. Accenture, Mastercard, and Visa are working together to bring the blueprint to clients and merchant networks, according to Digital Commerce 360. The shift is from an isolated prototype to infrastructure spread across global payment networks, and on those networks an error travels at the speed of payment.

Greetings to everyone from the table, Adam: the market has changed what it prices. Nscale, a young British AI infrastructure company, told investors a revenue figure close to 103 billion dollars after its deal with Anthropic, according to TechCrunch on September 4, 2026. That figure describes projections based on leases signed by clients, separate from current sales, as The Information clarifies. The consensus looks at data centers, GPUs, megawatts: it measures installed power and concludes that the advantage belongs to whoever owns more silicon. The consensus is looking in the wrong place. Computing performance falls along a cost curve, and whatever falls in price loses pricing power. The durable asset is the multi-year contract that locks in the client through duration and exclusivity. Value migrates from hardware to control of the commitments.

Good day to everyone, and yes Adam, history has seen it before. In 1949 the Western powers created COCOM, the committee that coordinated the technology embargo against the Soviet bloc. It operated until 1994 and slowed access to advanced machinery, and still, capital and expertise found alternative paths. In 1987 the Toshiba-Kongsberg case made it plain: machine tools reached Soviet shipyards by circumventing the controls, and submarine propulsion improved. Today Washington restricts hardware and leaves the flow of money untouched. The export controls on advanced chips to China were meant to slow Chinese artificial intelligence. Moonshot AI, a Beijing lab, trained its flagship model, Kimi K3, on a cluster of 20,000 Nvidia H200 accelerators. The context differs, the structure stays identical: when the money gets through, the silicon follows.

Welcome back everyone, and welcome back Adam: today's story starts with an architecture choice. In the second fiscal quarter of 2027, Salesforce moved the center of its product from screens to autonomous agents. For more than two decades the company had built its value on dashboards and workflows run by an operator in front of a screen. Agentforce, the new product, flips that logic. The agent becomes the operating unit: it qualifies leads, handles support requests, resolves tickets. The human task shrinks to strategic supervision. Whoever already owns structured customer data at global scale turns that data into autonomous actions almost immediately. That redesign around the agent, in place of the interface, explains most of the quarter's growth.