Meta's Muse Is the First AI Agent People Actually Use. That's Exactly What Scares Amazon.

There's a running joke in the agentic AI world that nobody actually wants an AI assistant — we all just like reading about them. Then Meta shipped Muse, and the joke aged poorly in about a week.

Muse is a consumer AI agent that does the boring adult stuff on your behalf. Books appointments. Shops. Keeps your calendar honest. It's powered by Meta's latest Muse Spark model, and it recently knocked ChatGPT off the top spot as the most-downloaded free app on Apple's App Store. For a company that spent two years losing the frontier-model conversation to OpenAI and Anthropic, that's a genuine win — the first one that feels earned rather than announced.

To be fair, the whole thing makes me a little uneasy. Muse works by getting access to accounts across your phone and then acting like a very fast, very forgettable personal assistant with your passwords. You're trading convenience for a piece of software that can spend your money. That's a real trade, and most people won't think about it until something goes sideways.

Amazon just drew the first line in the sand

Amazon didn't roll out the red carpet. It straight-up blocked Muse from accessing its store, citing its Conditions of Use and noting that Meta never asked permission. Meta, for its part, partnered with Shopify for shopping features and hasn't exactly apologized.

Look, I get both sides here. Amazon's ad business is built on humans scrolling its pages — an agent that skips straight to checkout bypasses the entire revenue loop. But this is bigger than one retailer's Terms of Service. Nikesh Arora from Palo Alto Networks put it well: every commerce, marketplace, or services app is going to have to decide whether to open APIs for consumer agents. The small players don't get a choice. Either the agent works for you, or the distribution gatekeeper takes a bigger cut.

And it's not like Amazon is a neutral referee here — it has its own shopping agent, Buy for Me, with brands opted in by default. Honesty check: that's not principled opposition, that's a turf war dressed up in legalese.

The ChatGPT crown is actually slipping

Comscore's AI Intelligence Report for Q2 2026 dropped this week, and the headline numbers are worth sitting with. ChatGPT still accounts for half of all prompt volume in the US — but its share fell 20% over the past six months. Gemini climbed to 30%, Claude to 11%. The margins aren't collapsing, but the trajectory is clear: people are splintering across assistants.

The quiet force here is AEO — answer engine optimization. Brands have realized that showing up in chatbot answers matters more than ranking on a search results page, and they're gaming all the assistants at once. Which means your visibility now depends on which assistant your readers actually use. Michael King over at iPullRank nailed it when he told publishers to stop asking "how do we show up in ChatGPT" and start asking "where is our audience asking questions." That's the kind of advice that's obvious in hindsight and ignored in practice.

I've been running the same queries across all three assistants for a few weeks now, and honestly? The answers differ a lot more than people assume. Gemini leans into terse utility, Claude writes like it's explaining to a friend, ChatGPT is somewhere in the middle. If you're building anything around AI discovery, treating them as interchangeable is a mistake that's about to get expensive.

The scary one: malware that votes on how to hurt you

Not everything this week is consumer-friendly. Researchers documented CLOSEDQUORUM, what they're calling the first Windows implant to use LLMs for its command-and-control. The thing queries up to four providers — Google Gemini, DeepSeek, Qwen, and Mistral — and lets the models pick from predefined post-compromise actions. Essentially, a small committee of AIs voting on the next step of an attack.

The obvious appeal for attackers: no hardcoded infrastructure, so it's harder to sinkhole. The obvious problem for everyone else: this is the "dual use" debate arriving in malware form, and it's not going to be the last example. Keep this in mind — the same public APIs you use to plan a trip are now being used to plan a breach. The security community is going to spend the next year chasing this pattern.

Europe finally getting serious about chips

On the hardware side, Germany and the Netherlands just committed €40 million over 20 months to fund small teams using AI to accelerate chip design. The Dutch launched NADI this month — their DARPA equivalent, with €500 million behind it — and it's teaming with Germany's SPRIND. The target: shave years off a design process that currently takes years, targeting "several orders of magnitude" of acceleration.

The most quotable line came from NADI co-founder Jelle Prins, who compared using NVIDIA chips for inference to "doing your shopping with a truck" instead of a grocery cart. It's a fair hit. The inference workload is where efficiency actually matters, and Europe's bet is that homegrown, AI-assisted design can produce chips that fit the job instead of brute-forcing it.

Will it work? No idea. DARPA-style agencies have a spotty record, and €40 million is pocket change next to what the US and China are spending. But it's the first time in a while Europe has a coherent story on silicon, and coherence alone puts it ahead of where it was a year ago.

The quiet backdrop

Worth a quick mention: OpenAI and Anthropic are pushing the UN for global risk-evaluation standards, and there's talk of a US-China AI alert system ahead of the Trump-Xi summit. And a Reuters/Ipsos poll found 73% of Americans think AI companies aren't doing enough to prevent disaster. All of this is governance noise, but it's the kind of noise that becomes policy before you notice.

Here's where I land. Muse proves agents can break through with normal people — if you give them something useful and stop talking about benchmarks. But it also proves the next big fight isn't model-versus-model. It's about who controls the access layer, who gets to be your representative online, and who pockets the margin. The model wars are basically settled. The agent wars just started.

---

Enjoyed this? You might like 7x24planning for practical planning tools.