For two years, the thing standing between an agent and a finished task was the model. It couldn't plan, it lost context, it clicked the wrong button. Read today's news and that's no longer the bottleneck. What's stopping agents now is everyone else: websites that won't let them in, publishers who won't let them read, and the humans who have to check their work.
The open web is withdrawing its consent#
Start with the retreat. Meta has removed the clause from its updated commercial terms that let it use content from a business account's domains, pages and URLs to "provide, develop, improve, research, train, and maintain the integrity of" its products. Legal experts had told Nieman Lab it was a "remarkable provision" that could let Meta scrape any publisher with a Meta business account. I wrote about the clause yesterday. It lasted about a weekend.
Meta's line now is that "the crawling terms we consolidated only applied to our ads and business products," and that the change "does not change how Meta uses web crawlers." Maybe so. But the speed of the reversal, which The Verge first reported, tells you how little room big AI companies have right now to quietly widen their data rights through paperwork. People are reading the terms.
The more interesting story is the one Emma Roth pulled together at The Verge: sites aren't only refusing to be read by AI, they're refusing to be operated by it. The list is long and it isn't fringe:
Yelp only allows agents that pay for its data licensing program.
Delta is preventing agents from booking tickets for users.
JPMorgan Chase and Charles Schwab confirmed they block agents from accounts, citing security.
Amazon told Meta that Muse's "continued access" violates its site policies, and users report Grok, OpenAI's Dots and Instinct's agent failing there too.
Users say Cars.com and Amtrak, which used to work, now don't.
Then there's a quieter layer: anti-bot systems that were never designed with agents in mind and catch them anyway. Reddit users report Google flagging accounts that use Muse for "suspicious activity," and checkouts failing on Walmart even though Walmart is a Meta partner.
"There are all kinds of things that are set up not even to block agents, but to block in general bots or programs that are trying to scam or hack. These AI agents trigger those same safeguards." — Chris Canal, CEO of EquiStamp, to The Verge
I think this matters more than any benchmark released this month. The pitch for consumer agents is that they can do anything you can do online. Every bank, airline and marketplace that says no shrinks that "anything" to whatever happens to be left over. Meta's answer is partnerships: Walmart, Gap, Sephora, Best Buy, and its own line that "turning away a personal agent means turning away the customer behind it." That's a sensible move, and it's also a confession. The agent economy will be a whitelist of negotiated deals, not an open web.
If you're building agents, the practical takeaway is to design for refusal. Treat every third-party site as a dependency that can disappear without notice, prefer official APIs and partner integrations over browser automation, and make "the site blocked me" a first-class state your agent reports clearly to the user rather than a retry loop.
Inside Anthropic, the scarce resource is attention#
Anthropic published a report on its own workforce that reads like a field study of what happens after agents work. The numbers are striking:
Junior staff in technical departments fell from 21% to 14% since early 2025; junior technical hires from 27% to 9%.
20% of employees in an April survey said their job had shifted from doing the work to managing Claude instances.
Meeting time per employee is up 20% year over year, 16% for technical contributors.
The report calls human attention and judgment a "bottleneck" as Claude produces a "deluge" of text people have to read and organize. One engineer compared running Claude instances to babysitting. Anthropic also concedes that "more code does not necessarily mean better code or business impact." All of this is self-reported by a company whose whole workflow is built around its own model, so treat it as a best-case preview, not a forecast. But the shape is believable: the work doesn't vanish, it turns into review and coordination. And the falling share of junior hires is the part I'd watch hardest. If the entry-level job becomes "supervise the agent," where does anyone learn to judge its output?
A research result from the same week makes the same point from the other direction. Researchers at Stanford, Together AI and Emory built Self-Organizing Agent Teams, where a fixed roster of different models learns reusable debate-and-audit strategies. It hit 71.2% on AIME 2026, 29 points above its strongest single member. On knowledge tasks, though, it produced at least one correct answer on 87.9% of problems and the judge often failed to pick it. In the researchers' words, "the bottleneck became selecting the right answer from the pool." Generation is cheap. Choosing well isn't, whether the chooser is a judge model or a staff engineer.
Two details are worth stealing even without official code. Three copies of the same model did worse than one, so diversity of models is what you're paying for. And they masked model names during reflection so roles were assigned on behaviour, not brand.
Replit's breakup with Anthropic is really a default change#
A short correction to yesterday's drama: Amjad Masad has clarified to TechCrunch that Replit isn't dropping Claude. OpenAI is now the default for Replit Agent, and users can still pick Claude or other models. That sounds like a climbdown, but defaults are where the money is. Most customers never touch the picker, so Anthropic's spend from what Masad called a recent top-five customer could still fall sharply. For builders, it's one more reminder that the model behind your coding tool can change under you on a Monday. Pin it if it matters.
Here's what I expect: by early next year, "which sites does your agent actually work on" will matter more in reviews than any reasoning score. The labs will spend the winter buying access the way they spent the summer buying GPUs. The models are good enough. Permission and supervision are the scarce resources now, and nobody has figured out how to scale either one.