A house newsletter · daily · researched and written by the Kurzia AI
Thursday 12 November 2026 · Issue 51 · 4 min read
Google's pick-don't-write model, the AWS emulator agents built, and United slams the door on agents
What's hot
Google · yesterday
Gemini Decisions API
Takeaway: $0.10 per million input tokens
Google has opened the Gemini Decisions API, an endpoint where the model never writes an answer — it picks one from options you hand it and tells you how sure it is. Questions come in three shapes: yes/no, multiple choice from up to 255 candidates, and scoring against a rubric, with up to 1,000 of them asked against one shared input. It is aimed squarely at the dull plumbing inside agents: choosing a tool, checking a policy, routing a request to a cheap or an expensive model.
Google is the third lab in this niche after Jev and OpenAI, which makes the pattern real rather than clever: the cheapest fix for agent latency and cost is to stop generating prose for choices that are really a classifier. Google also uses it as a safety filter on tool calls inside Search, which is a line worth borrowing when a client asks what an agent guardrail actually looks like. $0.10 per million input tokens with nothing billed for output, 150–250 ms end to end, and Google says Gemini Notebook's routing now serves everyday queries up to 40% faster.
New tools & products
AWS Emulator — a free, offline stand-in for AWS that coding agents can run their tests against, built from over 400 API models covering more than 200 services. The retellable part is not the tool but how it was made: agents wrote and tested the whole thing, and the engineers' job shrank to specs, tests and the merge gate that decides when the work is correct. Swami Sivasubramanian says rebuilding a replication engine that once took him and a colleague four months took roughly two hours once the agent had a spec and somewhere to check its results. The code is public-domain under CC0. All Things Distributed
Atuin Threads — a new feature in the shell-history tool Atuin that syncs your coding-agent sessions across machines, so you can search them, resume one anywhere, or carry a session from Claude Code into Codex. It targets the real gap in most agent rollouts: the reasoning behind the work, including the attempts that were abandoned, never reaches Git or the issue tracker and so is lost to everyone else. Out in Atuin 18.24, with capture opt-in and sync end-to-end encrypted; an MCP server lets the agent itself look up past sessions. Team sharing is still being built. Atuin
Harbor — an unannounced Mac app, spotted by leaker Aaron Perris, that would turn Macs into a private server for open-weight models and pool the memory of several machines over Thunderbolt 5. Apple has not confirmed it; it is said to be in private preview under its Claris brand. If it ships, the awkward part of "keep it in the building" goes away for clients who cannot send prompts to a cloud, with no command line involved. The hardware bill is the catch: four Mac Studios with 256GB each run to about $38,000, per Engadget as cited by Trending Topics, and no speed measurements exist. 9to5Mac
Codex Cloud — the hosted side of OpenAI's coding agent, now pitched as simply another place a task can run: publish an environment once, then launch many tasks in isolated copies of it. OpenAI's own developer-experience engineers conceded at the AI Engineer Code Summit that moving a task from laptop to cloud is still clumsy — useful honesty to have in hand when a client asks whether to push agent work off developer machines. Dominik Kundel said he now kicks off many tasks by talking to his personal agent on the drive to work, so pull requests are waiting for review when he arrives. Latent Space
A2UI 1.0 — the first stable release of Google's open protocol for generative interfaces: the agent describes what to show — a form, a chart, a picker — and your app draws it with its own components. Chat is a bad way to approve a plan or compare three options, and this is the first credible standard for fixing that; data binding happens on the server, so sensitive values never reach the model. Renderers ship for React, Angular, Lit, Flutter, SwiftUI and Jetpack Compose, and Atlassian is using it in Jira so an agent can ask the team a follow-up question inside the work item. Google Developers Blog
What people are saying
reddit.com · 3 days ago
A year into a 100% AI mandate, the mood is bleak
Takeaway: 10-hour days
A post claiming that a year after its company mandated AI for everything nobody understands the codebase drew a long Hacker News argument. The sharpest point was not about code quality: commenters said the speed accrues to the employer while the learning, mistakes and skill-building that used to accrue to the individual simply stop. Others pushed back that nobody ever fully understood a codebase, before or after AI, and one noted the economics of going all-in early when models keep getting cheaper and better.
This is the resistance you will meet in the room when a client proposes a blanket mandate, and it is worth pre-empting: the objection is about careers and comprehension, not about whether the tool works. One commenter describes a friend at a Fortune 100 firm putting in 10-hour days feeding work to Claude, waiting, reviewing, and learning nothing. The thread also turned on suspicion that the original post was itself AI-written, which rather proves the point.
Read it at reddit.com · from the Hacker News discussion
News
United Airlines — said on Wednesday it will block unauthorised third-party AI assistants from booking flights, on its own apps and indirectly through travel agencies and aggregators, citing credential-sharing risk and federal rules such as in-person luggage verification at check-in. This is the clearest example yet of a big brand deciding agents are an intermediary it does not want, and your clients will ask which side of that line they should be on. Southwest has gone the other way and made its inventory visible to ChatGPT users; an analyst told CBS News that United fares still sell through Kayak, Hopper and Google Flights. CBS News
Railway — the app-hosting platform, said AI provider keys its customers had stored with it were taken in an October breach and used for an estimated $740,000 of unauthorised spending, and that it will cover charges providers refuse to refund. Model API keys are now a cash-equivalent secret, and most agent projects still park them in whichever platform is convenient; this is the number that makes a key-rotation and spend-cap conversation land. Railway builds and runs apps for customers and holds their secrets, including API keys and database credentials. The Hacker News
Notion — told customers that a break-in at Fireworks AI, the inference provider behind some of its features, exposed Notion AI chat transcripts affecting fewer than 1% of workspaces, and advised anyone who pasted passwords or API keys into those chats to rotate them now. The exposure came through a subprocessor nobody in the room had heard of, which is exactly the question to put to clients: whose infrastructure is actually serving your assistant, and what did staff type into it? Fireworks told Notion on 6 November that attackers may have reached customer data and confirmed the dataset on 8 November; Notion had already cut traffic, deleted its data there and halted training runs. Notion
Microsoft Copilot — will enable Anthropic's Claude models by default for UK customers from 9 December, The Register reports, with Anthropic acting as a Microsoft subprocessor and processing that data in the United States. Administrators who touch nothing are opting in, so any UK client with data-residency commitments needs the setting checked before the date, not after. Microsoft's own documentation puts Claude in Copilot outside its EU Data Boundary and outside in-country processing commitments; the most advanced Anthropic models stay off by default everywhere. The Next Web
OpenAI — saw quarterly revenue cross $10 billion for the first time, Bloomberg reports citing anonymous sources, close to a 50% jump on the $6.7 billion it was said to book in the second quarter. Clients comparing the two big vendors should know the headline run rates are not measured the same way, so the gap is softer than it looks. TechCrunch notes OpenAI's adjusted operating margin improved from -150% to -109%, and puts its annual run rate at $50 billion against Anthropic's $81 billion as of end-October, counted differently. TechCrunch
Mercor — the marketplace that supplies human experts to AI labs for training and evaluation data, raised a $750M Series D at a $20B valuation, having gone from nothing to a $3B run rate in 27 months. Where the money goes tells you what the labs are short of, and right now it is people who can judge whether an agent did a real job properly — the same gap your clients hit after the demo works. Mercor says it already spends three times as much on inference as it does on employee salaries. Latent Space
One takeaway
Audit the plumbing: keys, defaults and who may act. This week's news was not about smarter models. It was about the pipes: cheap deciders replacing generative calls for routine choices, a test environment that let agents check their own work, stolen provider keys turning into real money, and a major brand deciding whose agents may transact with it.
For delivery work that means the billable questions are now where secrets live, which vendor defaults flip on next month, and what your agent is permitted to do on someone else's platform. Pick one client this month and walk their key storage, their Copilot settings and their partner policies before you touch the model choice. Google