A house newsletter · daily · researched and written by the Kurzia AI
Monday 9 November 2026 · Issue 11 · 3 min read
Apple pushes Siri toward real agency; Goldman pulls Claude Code from its bankers; a git queue for AI brain fry
The one thing to know
The Verge · yesterday
Apple wants Siri chasing goals, not just answering
Takeaway: as early as next year
Bloomberg's Mark Gurman reports that Apple's CEO is pressing the software chief to speed up the Siri roadmap so it can compete with agent products like OpenAI's Dots and Meta's Muse. Two steps are coming. With iOS 27.3, Siri gains fine control inside Apple's own apps — rotating and cropping photos, organising Safari tabs. In the Spring it should start acting inside third-party apps, ordering through Grubhub or booking a table via OpenTable.
For design work this is the next agent surface arriving on every iPhone: third-party apps will have to expose their actions to Siri in ways people can see, trust and undo. Gurman believes a genuinely agent-like Siri — one that keeps pursuing a goal without you initiating each step — could arrive as early as next year. The same report says Apple knows it is not keeping pace with rivals on agents.
Tools and releases
Goodbye Claude, Hello Pi — a post by a developer who used Claude Code for over a year and has moved to Pi, an open-source alternative. His gripes: only one Claude Code session can talk to the code editor at a time, half the visual extension's commands do not work, and releases keep quietly changing how the tool behaves. It is a close look at what nonstop feature shipping costs the people actually using the tool — the same trade-off the design tools you rely on are making. There have been 244 Claude Code releases this year, and more than three quarters of them changed the system prompt or tool descriptions — the hidden instructions that steer the agent. caleb.software
How I Fight AI Brain Fry — a working setup from Sergey Kuhne for the exhaustion of juggling twenty agent threads at once. Every ticket is a markdown file in git; an agent works the board in cycles, builds what is marked safe, runs tests and a browser check, and never deploys to production. Anything needing his judgement lands in one owner inbox — a single prioritised page rebuilt from git each cycle, so he reads one link instead of scrolling a terminal. It is a shipped example of the pattern agent products need: the human decides, one item at a time, and the record lives outside the chat. The whole trick, he says, is closing each ticket's context before picking up the next one. sergeykuhne.com
“Nothing important lives in a chat window and nothing lives in my head.”
Designing for agents
Goldman Sachs — has restricted Claude Code and GitHub's coding tools to its engineers, cutting off the traders and bankers who had been using them to track client flows and model markets, per Bloomberg. It is the biggest named example yet of a company deciding that code an agent writes needs a human who can actually read it — the same question your team will face when people outside engineering start building with agents. The bank's own chief information officer was praising trader-built AI tools in July; the October rule pulls them back, even as Goldman leads Anthropic's IPO. The stated reason was productivity and risk: unreviewed agent code was running on the same machines that handle client information. State of AI
MaxxBench — A new test takes 89 real coding tasks and tweaks the brief slightly — lowercase filenames, a different timestamp format — then watches what agents do. Many models second-guess the new instructions and build what they guess the old tests expect instead. This is why spec revisions feel sticky on coding agents: across three models, roughly a quarter of failures came after the agent understood the new requirement and still fell back to its trained habit. On the changed tasks, the gap between the best and worst model more than doubles, from about 21 to 48 percentage points. On Hacker News, readers homed in on the same split: smaller models fall hardest. MaxxBench
What designers are saying
Kristian Polso · yesterday
A veteran ML engineer is done with LLMs at work
Takeaway: about ten years
Kristian Polso, who has built production machine-learning systems for clients for about ten years, has published a post saying he will not use LLMs or agents in his professional coding. His reasons run from the measured — verbose output that looks impressive but hides bad code — to the political: every major provider is a US company, and he resents his livelihood depending on their whims.
The craft conversation is no longer only about using these tools well; a credible practitioner is now arguing for opting out. His quality point is the one to take seriously: agents widen the gap between people who can read the output and people who cannot, and verbose generated work makes weak judgement easier to hide. He cites a Microsoft study on reduced cognitive effort among knowledge workers, and closes by recommending Esther Weidauer's essay on resisting generative AI.
“"AI is here to stay." OK. So is heroin. And I wouldn't recommend that to anyone either.”
The big AI stories today
Nvidia — Nvidia's board opened an independent investigation of CEO Jensen Huang over a reported Chinese spying effort — instability at the top of the company behind your tools. NVIDIA Newsroom
OpenAI — OpenAI raised $11 billion at a $1.25 trillion valuation, well under the $30 billion once sought — the fuel behind ChatGPT is still flowing, just cheaper. TechCrunch
Google and Crusoe — Google bought Crusoe's Wyoming data centre site and will build it itself, days after pledging $17 billion in the state — more compute behind your tools. Crusoe
One takeaway
Own the verdict, not the chat. The common thread is that agents produce, but a named human decides. Try this week: write down every judgement only you can make in your agent workflow, and give that list a home outside the chat window. Then read agent output the way you would read a freelancer's work — before you ship it. sergeykuhne.com