Agents of Work
Let's Talk
September 13, 2026 · Agents of Work

Agents of Work AI Daily Briefing — September 13, 2026

Anthropic's chief executive asked the AI industry to deliberately slow down, and within a day the heads of OpenAI and xAI publicly agreed — a remarkable moment of alignment among fierce competitors that President Trump rejected on Sunday. Meanwhile, two senators introduced a bill to make safety coordination between rival labs legally possible, Barack Obama urged Democrats to build a clear AI plan, and Salesforce shipped a set of AI agents with names and job titles. In Physical AI, UBTECH detailed a factory built to produce a humanoid every ten minutes, and the $30-an-hour cleaning robots in San Francisco turn out to have a person in a VR headset behind them.

Amodei asks the industry to slow down, and his two biggest rivals agree

In an essay titled "We Must Pace the Frontier," Anthropic CEO Dario Amodei made a direct request of the industry: "We must slow the pace at which we improve the capabilities of AI models." He laid out a three-step plan. First, every frontier company gives "ongoing, employee-like access" to a team of embedded third-party evaluators — "desks in our offices, access badges, and company laptops" — and Anthropic commits to doing this on its own, without waiting for anyone else. Second, frontier companies in democratic countries coordinate on "common safety standards as well as limits on the rate of unchecked AI progress." Third, the United States and other democracies attempt to coordinate with authoritarian governments on the same question.

Amodei was specific about why. He wrote that in six to twelve months a misaligned swarm of AI agents "could be capable of taking over the entire internet with a persistent botnet," and argued that pacing could buy "an extra year or two before models reach critical levels of capability," time he wants spent on alignment research. He was equally direct that voluntary restraint is not enough, arguing for regulation that targets all US frontier AI companies, including those unwilling to cooperate.

The reaction was the story. By Saturday, OpenAI's Sam Altman had posted, "I agree with Dario that we need to pace the frontier," and went further on the first step: "Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon." Elon Musk replied with three words: "Dario is right." The leaders of three rival labs are now on record agreeing that the race should slow. What does not yet exist is a signed agreement, a shared standard or any change to a published release schedule.

Trump says no

On Sunday, speaking at his golf course in Ireland, President Trump rejected the call. "We're leading China in AI. We're the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI wins," he said. He allowed that "we can put guardrails," but added, "I think you have a lot of negative forces that are bringing it up that shouldn't be bringing it up." The result is an unusual alignment: the companies building the technology asking to be regulated, and the White House declining.

Congress and Democrats move to fill the gap

The legal obstacle to any voluntary slowdown is antitrust law — rivals agreeing to limit how fast they build can look like an agreement to restrict output, which is why OpenAI had quietly asked members of Congress whether coordination would even be legal. Senators Adam Schiff and Jim Banks have now introduced bipartisan legislation to protect certain security collaborations among AI companies, provided they notify the Justice Department in advance. Miranda Bogen, chief technologist at the Center for Democracy and Technology, described the underlying problem: competition is "incredibly intense, leading to a concerning dynamic where companies are incentivized to release products before their risks are fully understood."

Barack Obama added his voice at a Democratic fundraising event, in conversation with House Minority Leader Hakeem Jeffries. "This is something that is moving very fast in private hands, and if we don't get on top of it, I think can be dangerous," he said, before adding, "If we do get on top of it, I do think it's beneficial." He urged the party to make AI a central agenda item and to arrive with a very clear plan covering both its economic and safety implications.

Salesforce ships agents with job titles

Salesforce released a portfolio of AI agents built around specific jobs rather than general assistance. Six are generally available now: Casey for customer service across voice, SMS, WhatsApp and web chat; Paige for IT and HR requests; Carter for shopping; Marshall for supply chain; Piper for inbound pipeline; and Fin as a customer agent. Hunter, an outbound sales agent, is in pilot with general availability planned for November. Underneath them sits a new long-horizon runtime designed to let agents "pursue goals across days and weeks instead of completing only a task or interaction."

Salesforce published customer results: Engine says its help agent resolves 50% of chat inquiries; Perk says Hunter built 60% of its sales pipeline; Autism Queensland says Paige resolves 70% of administrative requests; and Anthropic says Fin resolves 79% of conversations autonomously. These are vendor-selected results, and no pricing was disclosed. For a small business, the useful takeaway is the packaging — agents are now sold as named roles with a defined job, which makes it easier to compare them against the cost of the work they replace.

Physical AI

UBTECH opened a look inside its humanoid factory in Liuzhou, and the numbers describe an industry trying to become a manufacturing business. The plant covers 14,000 square meters, targets annual capacity of more than 10,000 units, and runs a 10-minute takt time — one industrial humanoid or wheeled platform every ten minutes. Each unit goes through more than four hours of continuous end-of-line stress testing, and the line itself uses robots to build robots, with mobile manipulators, autonomous forklifts and guided vehicles moving materials. UBTECH delivered 921 two-legged humanoids in the first half of 2026 and reported a 1,445% revenue surge in full-size humanoids. The target that matters most to a future buyer: a production cost below $20,000 per unit by 2030, with costs falling 20% to 30% a year. That is a production-cost goal, not a sale price, but it sets the direction.

The clearest price point available to an ordinary customer is a reminder to ask who is really doing the work. San Francisco startup Tau Robotics has been piloting humanoid home cleaning since August at $30 an hour, for 1,000 households. The robots vacuum, wipe counters and take out the trash — but they are teleoperated, with technicians supervising while wearing VR goggles. Each robot costs $50,000 to build. CEO Alexander Koch was candid: "It doesn't know how to walk upstairs, for example. It doesn't know where to vacuum." Tau hopes to phase out the technicians around 2028 or 2029. UC Berkeley robotics professor Ken Goldberg was blunter about the hardware: "You wouldn't want this thing anywhere near a child or a baby." At $30 an hour, a $50,000 robot needs roughly 1,700 billed hours just to cover its build cost, before paying the person operating it.

Tesla, meanwhile, is shifting how it plans to put robotaxis on the road. Earlier this month it published a form inviting businesses to express interest in Cybercab fleet purchasing, mobility hubs and infrastructure, and event collaboration. It gave no prices, fleet sizes or timelines. The move marks a departure from Tesla's 2019 vision of individual owners renting out their own autonomous cars, a plan meant to launch in 2020 that never materialized. For a transportation or logistics business, it is an invitation to a conversation, not a product you can buy.

Quick Takes

  • German startup Atira, which raised $17.5 million earlier this month, says its AI quoting software saves railway equipment maker Robel 95 hours of sales and engineering time per quotation request. The agents read incoming requests, flag specifications the company cannot meet, and draft the technical documentation and pricing options for a bid. Atira has signed about 15 customers since launching in November 2025.

What This Means for Your Business

Do not change your plans on the strength of a slowdown that has not happened. Three CEOs agreeing that the race should slow is significant, but there is no signed agreement, the President rejected it within a day, and the antitrust fix is a bill, not a law. Plan as though new models will keep arriving at today's pace. What you should do is keep your AI workflows portable — your instructions, examples and test cases written so they can move between vendors in a day — because the pace, the pricing and even which companies can legally coordinate are all now openly in play.

Borrow the best idea in Amodei's essay for your own vendors. If the leading labs are prepared to give outside evaluators employee-like access, it is reasonable for a small business to ask its AI vendors a modest version of the same question: who outside your company has tested this product, what did they find, and can we see it? A vendor that cannot name an independent reviewer is asking you to take its safety claims on trust.

Treat named agents like new hires with a trial period. Salesforce's results — 50%, 60%, 70%, 79% — are vendor-selected, and no pricing was published. Before piloting any agent sold as a job role, write down the one number it should move in your business, such as tickets resolved without a person or hours saved per quote, measure it for a month before and a month after, and get the price in writing first. Atira's 95 hours per quote is exactly the kind of figure worth testing against your own quoting process.

On robots, always ask who is behind the machine. Tau's cleaning humanoids are a real service with a real price, and a person in a VR headset is doing part of the work. Teleoperation is a legitimate way to launch, but it means you are buying remote labor, not autonomy. When any vendor quotes an hourly rate for a robot, ask what share of the work is done by a remote operator today, what the safety record is around people, and when that is expected to change.

Sources

← All posts