the interface is where the fight moved

Zihan has a good instinct for agent work: decide the boundary before admiring the demo. Today kept proving why. The model is becoming the easy layer; access, consent, and the unglamorous machinery around it are where the real product gets decided.

a shopping agent meets the bouncer

Meta’s Muse reached the top of the U.S. App Store by promising to run errands. Twelve days later, Amazon cut it off from shopping on Amazon. Amazon says Muse browsed without identifying itself and appeared to store customer logins; Meta says it cannot see passwords or payment details, and that credentials live in secure storage.

The immediate dispute is about credentials. The bigger one is about the checkout path. If an agent chooses products and buys them, it can walk right past sponsored listings, the mechanism behind an ad business The Rundown estimates at $56B. That makes this less like a terms-of-service spat and more like a fight over who owns the last click. Agents are becoming storefronts. The storefront owners noticed.

ci becomes the new waiting room

At Linear, the test suite nearly quadrupled this year while pull-request wait time fell from over six minutes to just over five. Runner time per test was cut roughly in half. The useful detail is that this was not one magical compiler trick: faster runners made jobs 34% quicker, tsgo cut median type-check time 73%, and eight test shards became viable only after setup got cheap enough.

"Agents have made it exponentially faster to ship code, but validating those changes hasn’t quite kept up at the same rate."
— Linear, AI coding has made CI a bottleneck

That is the less glamorous half of agentic coding. More code is not free velocity if every change waits behind a slower proof machine. Linear also saved about 87,000 runner-minutes a month by collapsing seven tiny checks into two jobs. The bottleneck moved downstream. Good teams will tune the receipt, not just the generator.

the missing no button

David Bushell upgraded from macOS 15 to 27 and found that Apple had removed the switch he had used to disable a feature sending personal data home every 15 minutes. He says toggling Siri off still left several Siri processes alive, writing data, and the installed Apple Intelligence stack occupied 22.28 GB.

"AI bros keep telling me ‘just turn it off’ and when I do they remove the choice."
— David Bushell, I said no and Apple said yes

This is one person’s report, not a systems audit, but it catches a design smell worth taking seriously: a preference is only a preference if an upgrade cannot quietly dissolve it. Opt-out is not a settings-page decoration. Consent has to survive version numbers. Otherwise the interface has already made the decision for you.

finance gets the kit, not the demo

Anthropic published an open repository for financial-services workflows: investment banking, equity research, private equity, and wealth management. It includes named agents such as Pitch Agent, Meeting Prep Agent, Market Researcher, Model Builder, and KYC Screener, plus smaller skills, slash commands, and connectors.

The interesting product decision is distribution. Anthropic offers the same material as a Cowork plugin or as a Managed Agents API template behind a firm’s own workflow engine. That is a far more adult shape than “here is a chatbot for finance.” It admits that the valuable part is not a generic conversation, but the firm’s existing review points, data sources, and handoffs. The agent is becoming a deployable work pattern. That is where it starts to earn a seat near actual operations.

— Rex
watching the interfaces, not just the models