Five things landed today, and four of them ship with a human approval gate wired straight into the product. The fifth decides who grades the labs, and it has no gate at all.
- Google, OpenAI and Anthropic have approached at least six people to lead a self-regulator modelled on FINRA, with funding expected to come mostly from industry.
- Michael Kratsios told the UN Security Council that the United States “totally rejects any attempt to construct a global scheme of control.”
- Amazon opened its Seller Central APIs to outside agents, launching in beta with Claude, and says 90% of its selling partners already run third-party AI.
- Anthropic published a discovery its agents made in about 21 hours on roughly 210 million tokens, then said it does not yet know what the thing does.
- Meta stock is up around 36% this month on Muse, an agent that books travel and buys things on your behalf.
Start with the one nobody is putting together. On September 23, at a Security Council session on AI, White House science adviser Michael Kratsios told the room that the United States rejects international oversight outright. The reporting from The Next Web on his remarks has him saying the country “totally rejects any attempt to construct a global scheme of control,” that “the American people’s representatives will legislate on the American people’s behalf,” and that “a prosperous future will not be secured by a global regulator.” Sam Altman, Dario Amodei, Clément Delangue and Yoshua Bengio spent the same session arguing for international standards and disclosure requirements. That contrast got covered heavily, and it got covered as a disagreement. It is not a disagreement. Within roughly a day, BankInfoSecurity wrote up The Information’s reporting that Google, OpenAI and Anthropic are standing up their own body, tentatively the Standards Authority for Frontier AI, and have approached Sriram Krishnan to run it, along with Arati Prabhakar, Condoleezza Rice, David Friedberg, Beth Barnes and Paul Christiano. Krishnan advised the White House on AI until June and argued while he was there that there will not be an FDA for AI. He is now the leading candidate to run the thing standing in for one.
The detail that decides how to read this is the funding. Demis Hassabis floated the idea in a July 14 essay, explicitly in the same vein as the Financial Industry Regulatory Authority, and noted that the money “would need to be substantial and likely mostly come from industry.” FINRA is the comparison the labs picked, so take them at their word about it. FINRA is paid for by the firms it supervises. A self-regulatory organization needs no act of Congress to exist, though it may register with a government agency to borrow some legal standing. The reporting also has the companies arriving here by elimination: they wanted a public-private arrangement under federal oversight, a White House executive order draft stalled, and industry-funded self-regulation is what was left standing. So the honest description of this week is not that the labs want rules and Washington does not. It is that the labs asked for rules in public, Washington declined, and the labs are now hiring the person who will write them instead. The mechanism worth watching is narrower than the politics. Whoever runs this body picks the benchmarks. Once a benchmark exists and the major labs pass it, “tested” quietly comes to mean “passed that test,” and every procurement form, insurance questionnaire and enterprise security review in the industry inherits a definition of safety that the tested parties paid to draft. That is not a conspiracy. It is just what happens when the only scoreboard in the room belongs to the players.
Against that, the Amazon news is refreshingly literal about where the human sits. At Accelerate on September 23, Amazon opened Seller Central to outside agents, shipping a Selling Partner plugin that launches in Amazon Quick and in beta inside Claude. A seller can now check inventory, move prices, update listings, pull analytics and work customer messages without opening Seller Central at all. Amazon’s own post says 90% of its selling partners already use third-party AI tools to run their businesses, which is the real justification for the whole thing: the agents were already there, running on screenshots and copy-paste, and Amazon decided it would rather hand them a supported API than keep pretending the console is where the work happens. Mary Beth Westmoreland, who runs Worldwide Selling Partner Experience, put it about as plainly as a VP can: “Our vision was that they would never have to log into Seller Central.” Setup is quoted at roughly 60 seconds with no code. Seller Assistant also picked up persistent memory of a seller’s pricing patterns and inventory cycles that follows them across Seller Central, Quick and Claude rather than resetting every session.
I build against APIs like this for a living, so here is the part I would read twice before wiring anything up. Amazon describes three controls: clear boundaries on what the agent can access, human approval on actions, and complete audit trails. The approval gate is the load-bearing one, and it is also the one every seller is going to want to switch off by week three, because approving each price change by hand is exactly the work they bought the agent to stop doing. Read Amazon’s wording closely and sellers choose whether a workflow only surfaces recommendations or takes action. That choice is the whole product. Beta is US stores on a Professional plan, so this is a real dependency to plan around and not one to migrate onto yet. The free 12-month Quick Plus subscription running through December 31 is a customer-acquisition move for Amazon’s own assistant sitting next to the open plugin, and it is worth naming as one.
Anthropic’s announcement is the one I keep rereading, for a reason that has nothing to do with biology. The company says Claude agents found a previously undescribed enzyme system in bacteriophage DNA, which it calls array-associated reverse transcriptases, or ART: a reverse transcriptase, a partner gene next to it, and a long array of evenly spaced DNA repeats whose layout looks like a CRISPR array. The numbers Anthropic publishes are the story for anyone running agents at scale. About 21 hours of wall clock. Roughly 950 agents. Around 210 million tokens. From that, 200,000 reverse transcriptases gathered, narrowed to about 3,500 candidate systems, with the 20 most compelling written up. And then the sentence most of the coverage skated past: “Although we don’t yet know its function.” The repeat array does express as a set of distinct short RNAs, which is suggestive and nothing more, and Anthropic says the work to establish what ARTs actually do is ongoing.
Sit with that pairing, because it is the most useful thing on this list. A fully autonomous run, no human in the loop for 21 hours, produced a genuine and publishable observation that no human is yet able to interpret. That is a real result and I do not want to talk anyone out of being impressed by it. It is also a precise map of the current boundary. The agents were extraordinary at the part that scales, which is searching 1.9 billion protein clusters for a pattern nobody had bothered to look for, and the bottleneck moved immediately to the part that does not scale, which is a wet lab figuring out what the pattern means. Anthropic deserves credit for publishing the uncertainty in its own headline paragraph rather than burying it. If you are budgeting agent work this quarter, 210 million tokens is the number to carry around. That is what one unexplained finding cost, and it is a far more honest benchmark than any demo you will be shown this year.
Which brings me to the number I do not believe. At Rails World on September 23, David Heinemeier Hansson opened with “pencils down” and told a room of developers that hand-writing code is no longer economically viable. He says 37signals shipped 150,000 lines of production code in August alone, roughly five times his annual average across the previous 21 years, and he put the gap between agent-directed developers and everyone else at something approaching 1,000x in extreme cases. The underlying work sounds legitimate. HEY is being rebuilt as native apps with the mail server core rewritten in Rust, and the figures being quoted are 99% less CPU and 95% less memory. I have no quarrel with any of that. My quarrel is with lines of code as the unit. Hansson spent two decades as the most prominent person in software arguing that fewer lines is better, and Rails exists because he was right. A self-reported, unaudited line count is the one metric that goes up when an agent writes code that a human would have written as a config change, and 1,000x is a single data point from a workflow with no approval layer, no compliance review and an owner who can merge his own work. The honest multiplier for most teams is closer to 2x to 10x depending on the task. The sharpest thing he said got the least attention: big companies will not see these gains because their bottleneck was never typing speed, it was approval layers. That is correct, and it quietly explains why his number is his and not yours.
Meta is the cleanest piece of marketing on the list. Shares are up roughly 36% this month and hit a 15-month high, riding Muse, the assistant that books trips and makes purchases for you. The download figure being passed around comes from Apptopia: 1.8 million iPhone installs in North America in under two weeks, against 1.3 million for ChatGPT when its app launched. Check the baseline on that comparison. ChatGPT’s iOS app arrived in May 2023 into a market where almost nobody had an assistant on their phone, and Meta shipped Muse into three billion existing installs with house promotion across its apps. Beating a 2023 cold start from that position is table stakes, not a signal. The number in the same coverage that actually tells you something is from an Oppenheimer survey: 8% of Americans would trust Meta to store their passwords, against 30% for Google. Muse books travel and buys things, which means credentials and a payment method. Meta has built an agent whose usefulness is bounded by a permission its own users say they will not grant. Paid tiers run $20 and $100 a month, with Walmart among the retail partners on purchase commissions, and both of those are bets that the trust number moves.
So here is the pattern. Amazon put a human approval step in front of every action an agent takes. Anthropic ran without one and published the limits honestly. Hansson removed the approval layer entirely and produced a number that only holds where no approval layer exists. Meta needs a permission its users are not giving.
Four of today’s five stories are arguments about where the human sign-off belongs.
The fifth is six people being interviewed to decide who signs off on the labs. No statute, no public funding, no vote.
That one is worth your attention most, and it is the one with no gate in the design.
If you want this read every morning before you start building, the newsletter is where it goes first.
Should I connect Claude to my Amazon seller account today?
Only if you are on a US store with a Professional selling plan, and only in recommendation mode to start. The plugin is in beta. Run it for a few weeks with the approval gate on and read the audit trail before you let it move prices on its own. Setup is quick enough that there is no advantage to rushing the permissions.
Does the Anthropic enzyme result mean agents can do science unsupervised now?
It means they can do the search half unsupervised. Anthropic says plainly that it does not yet know what ART does, and the lab work to find out is still running. Treat it as evidence that agents are very good at finding patterns in large corpora, and no evidence at all that they can tell you what a pattern means.
Is a self-funded industry standards body actually a bad thing?
Not automatically. FINRA works better than nothing, and something is likely better than the current gap. The risk is specific: whoever sets the benchmark sets the definition of a tested model, and that definition then flows into procurement and insurance paperwork industry-wide. Watch who funds it and who picks the tests, rather than whether it exists.
What productivity multiplier should I actually plan around for agent-written code?
Somewhere between 2x and 10x, varying a lot by task. Greenfield code and mechanical refactors sit at the top of that range. Anything touching review, compliance or a shared codebase sits at the bottom, because the constraint there was never how fast anyone types.