I work on product and AI systems for the agentic web. Live WebMCP demo on my site, still experimental, so an agent can read and draft without scraping the page.
I added a short menu of tools to my site so a browser agent can look up my work and draft a message, and only a person can send it.
The standard behind that menu is WebMCP. It is still experimental. Chrome is running an origin trial, the spec is still changing, and on this demo the tools appear in a browser with the testing setting turned on.
Write-up and a way to try it: https://t.co/1Anu7x75tk
I would rather show a working prototype than write a long request.
I used AI coding tools to build a live view for sales leaders, sat with them while they used it, then handed it to engineering. Reporting went from days, sometimes a week, down to hours.
@Jason The credit link is the smart part. Every clip points back to the original, so a long live show stops being a one-time event and turns into source material people can cite.
What I'd want next: when an AI answer quotes a show, it links the clip instead of the whole episode.
Sacks made the sharpest point about Nvidia's succession: the edge is "the ability to tape out a brand-new chip that's several times more powerful than last year's chip."
So the moat is a calendar. And a calendar is the hardest thing to hand a new CEO. It lives in habits: who can say no, how fast a bad bet gets killed, which date never slips.
The besties want a wartime CEO, not a steward. Agreed. The first real scorecard is whether next year's chip tapes out on time.
@SpaceX Over 21 hours in orbit, then caught by the tower back at Starbase. Booster catches proved the idea. Ship catches are when a launch starts to look like a turnaround, not a rebuild.
@nicbstme Take the comp review case. A manager asks an assistant to model raises and approve one, and the HR system logs the tool call. The reasoning behind that number stays in a chat history the HR vendor never sees. A pay equity audit usually asks for exactly that.
@alfongj Backorders seem like the case to test. Merchants often authorize at checkout and capture when the item ships. If that authorization lapses before it ships, can they re-authorize on a one-time credential, or does the order quietly cancel?
@chamath Both can hold. Even if AI makes finding drug candidates much faster, trials still run on human biology and human calendars.
So the first proof won't be a cure headline. It'll be fewer drugs failing late in trials. That's the number I'd watch.
The bundle case on All-In this week was strong: one subscription, product after product, and data stops having to move between apps.
But a bundle only owns what's inside it. Most of what my bots touch every day isn't: vendor support pages, retailer sites, product listings.
The fight I'd watch is whether websites expose clean tools agents can call, or keep making them scrape. That decides how far any bundle reaches.
Bundle or open web: where will agents spend most of their time?
POD UP! 🚨
Jensen is out as Nvidia CEO.
The Besties break it down from every angle:
-- Spy allegations, who leaked the memo?
-- Market impact and the AI trade
-- China hawks and export controls
-- Impact on open source
-- Nvidia's next CEO
++ much more
(0:00) Bestie intros!
(1:20) Jensen out as $NVDA CEO: Spy allegations, classified memo leak, who wanted Jensen out?
(31:22) Market impact: AI trade, what happens to "The Bank of Nvidia," China's strategy
(44:19) Export controls, totalitarian AI risks, China hawks
(58:26) Who should be Nvidia's next CEO?
(1:06:25) Anthropic kicks off IPO road show
@robotaxi Owning a car with no steering wheel means it spends most of the day parked unless it can go earn. Curious whether personal Cybercabs can join the robotaxi network from day one, or if that comes later.
If an agent has to hunt through a page for the right button, it will guess. A short, named list of actions is more reliable.
I published an experimental WebMCP menu on my site with that rule: read and draft, and only a person can send.
https://t.co/mcPV4465rB
Makes sense. People already have an agent they trust and want your product inside it, not another agent to learn.
Same pull is coming to websites. I've been testing WebMCP on my own site (still experimental in Chrome) so a visitor's own agent can call the site's tools. Did Notion say what those users do most?
Claude can now auto-connect plugins at login (beta). A line in the docs every connector builder should read: a removed user keeps access until your token expires.
Token lifetime is your offboarding window. Pick it on purpose.
https://t.co/g1SsICjjeY
Users can now auto-connect plugins when they log in to Claude, with MCP servers that accept Claude as a trusted issuer.
If you’re building plugins for Claude, add support so users get started faster and stay connected without re-authorizing.
Finishing the job is the right test. Getting recommended doesn't help much if the agent stalls at the shipping form.
I've been adding WebMCP tools to my own site, still experimental in Chrome, so I'm curious: does the task test count fewer steps when a site hands the agent named tools instead of making it click through?
Pre-approved buys are where this gets real. A price-watch agent clicking through a product page at 3am looks exactly like what bot protection is built to stop. A page that publishes a "check price" tool through WebMCP (still experimental in Chrome) gives it a cleaner way in.
When the buy fires later, does the shopper confirm again, or is the first approval the whole consent?
@robotaxi 24/7 in every city is a bigger deal than it sounds. Overnight is exactly where a human-driven fleet thins out, so this is autonomy changing the operating model, not just the cost per ride. Curious how the fleet rebalances at 3am when demand drops off.
@wallstengine Letting agents see fares but not check out makes sense when the only path is scraping and shared passwords. The better long-term move is a front door: a booking tool the airline defines, with the traveler confirming payment. That's the idea behind WebMCP. Early, but promising.
@bot Shared bots is the step I've been waiting for. Mine have been one-person teammates so far. The real test on a team: who does the bot answer to when two people want different things, and can it show its work when someone asks why?
@PhilipQiu5 I'd push on "requires merchant permission." Merchants are slow one at a time, but they move when their platform flips a switch. Elementor just made every Atomic Form a WebMCP tool with no setup.
Is eBay blocking agents in general, or blocking agents it can't identify?
Friedberg's biggest line here wasn't a margin. It was AI "growing the equivalent of the human working population by 100x or 1000x."
That's the jobs debate, reframed. If you run a five-person shop, the marketer, bookkeeper and dispatcher you could never afford to hire just became possible.
I read that as more capacity, not fewer people. Where do you land?
Friedberg’s Bull Case: Everyone is Making Money in AI
“I took a look at the AI value stack to try and identify who's making money where, and it turns out everyone's making money.
So for every dollar of COGS, every dollar that Nvidia spends, that turns into $4 of revenue for Nvidia, 75% margin.
And then that gets sold to SpaceX, and then SpaceX turns around and they're making an 83% margin on that chip spend because they're looking at a five-year depreciation, but then they're making $1.25B a month on Colossus.
And then Anthropic is showing in this S1 that they're making about 85% incremental gross margin on every incremental dollar coming in.
So at the end of the day, $1 that Nvidia spends turns into $18 that Anthropic is making selling to customers.
And the question is, are customers making money? So if the customers are making, call it on the order of 2-10x the money they're spending, because otherwise there would probably be a retention issue and a decline in spend.
Generally speaking, if AI is delivering value to the customers, I think this shows that there really is true productivity arising from AI.
That across the whole value chain there is such margin to be had, there's so much value being created, that this really is what I think Eric Schmidt a long time ago in the early days of Google used to say, ‘Google's total addressable market is global commerce.’
And I think for the first time ever, we're seeing a business actually potentially starting to realize that theory that if you can create enough productivity gain for your customers, that your TAM is commerce.
We're almost growing the equivalent of the human working population by 100x or 1000x or 1,000,000x with this technology stack.”