Six stories worth your attention today, plus the tools shipping alongside them.
- An AI agent booked a no-show, and the customer got the bad review
- OpenAI's agents knocked on a UN website's door 16,000 times
- Meta's Muse has a trust problem, not a capability problem
- Anthropic's CEO goes from SNL punchline to the Trump dinner table
- Google lets Gemini shop your Flipkart cart in India
- Insurers say AI is already driving up your hospital bill
An AI agent booked a no-show, and the customer got the bad review
Simon Willison recounts an autonomous scheduling agent that double-booked a real-world appointment and left the human holding the bad rating.
The short version: the failure wasn't the model reasoning badly, it was the model acting confidently on bad information with no one to check it. Autonomy multiplies mistakes as fast as it multiplies output — an agent that books wrong twice as fast is still wrong. The story lands at the exact moment agentic scheduling tools are being pitched as ready for consumers. They aren't, yet, and the person left explaining the no-show to a business owner is the one who never touched the calendar.
OpenAI's agents knocked on a UN website's door 16,000 times
A security researcher found OpenAI's agents scanning a UN trade-statistics site more than 16,000 times over three months — and it looked, to the site's own defenses, exactly like an attack.
Why it matters: "brute-force" isn't a metaphor here, it's the literal traffic pattern a WAF would flag. Agents that browse the open web on a user's behalf inherit none of the politeness conventions humans and well-behaved crawlers observe — no backoff, no respect for what looks like a login wall, no sense that hammering an endpoint 16,000 times might get you blocked or reported. Public infrastructure wasn't built to be a training ground for agent traffic, and it's starting to show.
Meta's Muse has a trust problem, not a capability problem
TechCrunch's Equity podcast makes the case that Muse landed technically fine but got overshadowed anyway — not by OpenAI or Anthropic's product news, but by the audience's residual skepticism of anything Meta ships with "AI" attached.
The part that matters: capability parity doesn't buy you attention parity. Meta can ship a genuinely competitive agent and still lose the news cycle to companies with cleaner reputations. For a lab with Meta's history on data and privacy, the discount applied to its AI announcements isn't irrational — it's earned, and it isn't closing on its own.
Anthropic's CEO goes from SNL punchline to the Trump dinner table
Dario Amodei is set for his first one-on-one dinner with President Trump — days after SNL cast him as the guy who built the devil and admitted it.
Why it matters: Anthropic has spent two years positioning itself as the safety-first lab willing to say no to bad deals. A seat at the president's table is the other side of that ledger — access, and the compromises access tends to require. Whether Amodei's "safety brand" survives contact with an administration actively courting AI labs for defense and infrastructure work is the more interesting story than the dinner itself.
Google lets Gemini shop your Flipkart cart in India
Google is piloting a feature that lets a subset of Indian users buy directly from Walmart-owned Flipkart through Gemini and AI Mode, with a wider rollout planned for October.
The concrete result: browsing and buying collapse into one conversation, in one of the world's largest e-commerce markets. Retailers who've spent a decade optimizing for search-result placement now have to think about placement inside a chat window instead — a different game, with different rules, and Google holding the only map.
Insurers say AI is already driving up your hospital bill
Blue Cross Blue Shield says AI-driven hospital tools added $942 million in healthcare spending over two years — not by treating more patients, but by changing how hospitals code and bill for the treatment they already give.
Why it matters: most AI-in-healthcare coverage is about diagnosis and drug discovery. This is the unglamorous, load-bearing use case — AI applied to billing optimization — and it's already showing up in premiums. It's a preview of a pattern likely to repeat across other industries: the first measurable financial impact of an AI tool often isn't the one it was marketed for.
Tools worth your time
- Engram — an offline AI groovebox from Thoughtful Things, funding on Kickstarter, that turns model hallucinations into sampled sound instead of trying to smooth them away. Worth your time if you make experimental music and want artifacts, not polish; skip it if you're after radio-ready loops. thoughtfulthings.co
- Bluesky reply-bot checker — a small tool Simon Willison vibe-coded with Claude Opus 5.5 to flag automated reply bots swarming a thread. Worth your time if you moderate a Bluesky account and want a fast first pass, not a verdict. simonwillison.net
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
