A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 11 November 2026 · Issue 13 · 4 min read
Amazon's 98% cost cut, Cursor Projects goes team-wide, and Notion's exposed AI chats
The one thing to know
Strands Agents (AWS) · yesterday
Amazon cut a ticket-labeling bill 98% by dropping an LLM
Takeaway: roughly 44 times cheaper
Two Amazon engineers replaced Claude Sonnet 4.6 — labelling tickets across one of the company's largest internal systems — with a much smaller 'decision model': one that cannot write text, only return a probability for each label in a fixed catalog. Quality matched production Sonnet at 98% lower cost.
Plenty of teams run agents doing exactly this kind of work — triage, routing, classification — on a frontier model with a long prompt. The copyable part isn't the model but the migration: run both on live traffic, review every disagreement, keep a switch that reverts in under a minute. Per million tickets, Sonnet cost about $22,200 and the decision model about $500 — roughly 44 times cheaper — and average response time fell from 4,215 ms to 897 ms. The engineers are clear about the limits: it only works when the task is a closed choice that fits in a few thousand tokens.
Read it at Strands Agents (AWS)
“Running both on real traffic, then judging every disagreement, told us more than any offline eval would have.”
How teams are doing it
LeadDev's three questions — A LeadDev piece gathers DX, AWS and DORA researchers on what leaders should measure before 2027: developer wellbeing, accountability for agent-built work, and delivery health. If you're being asked to prove agent ROI, the useful frame is here — most governance is written down, almost none of it is enforced. It cites 41% of developers showing signs of AI-coding compulsion, 60% of agentic spend going to verification (McKinsey), and under a quarter of orgs with AI policies having actual controls. LeadDev
“A policy document states what you intend to happen. A deployment gate actually makes it happen.”
A trade-secret suit names the coding agents. Point72's quantitative arm Cubist has sued a dismissed senior researcher, alleging he lifted proprietary source code on his way to a rival firm and that Anthropic's Claude and OpenAI's Codex were the tools used. This looks like the first high-profile trade-secret case built on everyday coding agents — expect legal or compliance to ask what your policy actually says about pasting company code into third-party AI tools. Cubist is the quantitative trading unit of Point72 Asset Management. Techmeme (Bloomberg)
Railway — Cloud host Railway says attackers exploited flaws in its infrastructure in October to steal customers' secret credentials, and it believes they were after AI provider keys to resell. Model API keys are spendable cash; if your team keeps provider keys in hosting or CI platforms, this is the week to add spending caps and a rotation routine. Unauthorized spend on the stolen keys is estimated at about $740,000. Techmeme (BleepingComputer)
Lenny's Newsletter — Lenny Rachitsky's analysis of the late-2026 tech job market finds product managers benefiting most from AI, demand for engineers plateauing, and designers losing ground. Relevant to your next headcount pitch — the easy growth story for engineering roles is no longer automatic. Its sharpest line: the biggest AI hiring boom isn't in software. Techmeme (Lenny's Newsletter)
Agents and plans
Cursor Projects — Cursor's Projects workspace — until now a way for one person to direct a fleet of agents — now takes whole teams: shared docs and specs on top of private chats, a task board showing which person and which agent is on what, and agents that talk across teammates. Agent coordination becomes something a whole team can standardise on rather than a solo habit; on-call hand-off is the obvious first use. It's rolling out in beta to all Cursor users, and several Cursor teams already run their on-call rotations inside a shared Project. Cursor
“Multi-agent systems have so far served one person. Projects is a multi-agent system built for a whole team.”
Claude interactive connectors — Anthropic says Claude can now open and work inside full enterprise applications — Adobe, Atlassian and HubSpot first — editing a Confluence page or an Adobe design from within the chat, with a person approving each change. If Claude is in your stack, non-engineers will start driving your other tools through it; existing app permissions carry over, and admins can restrict individual capabilities. Anthropic reports Mercado Libre has Claude in front of more than 26,000 employees, 78% using it weekly — the company's own figures, not independently verified. The New Stack
“For years, we've learned to work around software. Now Claude works around you.”
Notion security notice — Notion says a breach at Fireworks AI, one of its model providers, may have exposed Notion AI chat transcripts in fewer than 1% of workspaces; the exposed chats carried no author, workspace or tool results. Notion has cut Fireworks off, deleted its data there and halted training runs. AI features often run on third-party inference vendors nobody in your org vetted — this is the incident to show your security team when asking who serves your vendors' models. Owners of affected workspaces can review and download exactly what was exposed, inside Notion itself. Notion
“We recommend that you treat this data as if it were exposed.”
Tools around the agents
AlphaProof Code — Google DeepMind open-sourced AlphaProof Code, a framework where AI agents produce formal verification — machine-checked mathematical proof that code does what it claims. Pointed at Signal's messaging protocol, the agents wrote six million lines of proof in four weeks, and humans reviewed only the one percent the agents flagged. The strongest signal yet that 'who verifies agent-written code' gets answered by proof rather than more review — think critical components, not everyday features. Two real, previously fixed Signal bugs were slipped back in as a test; the system proved both insecure. The proving agents ran on Gemini 4 Carbon, which went fully available the same day at an introductory $2 per million input tokens. Google DeepMind
What people are saying
Evan X. Merz — An engineer laid off from Meta — who now generates 100% of his code with AI at his new employer because that's the expectation — argues the result is brittle: shallow tests, needless complexity, fading human understanding, and pull requests arriving faster than anyone can properly review. What's new is who's saying it: not AI resisters, but people following the mandate. His warning signs are checkable — rising incident counts, longer recovery times. He writes that at Meta, 90% AI-generated code was 'the minimum for a decent rating' in his last months there. Evan X. Merz
“AI codebases are software made by AI for AI.”
The big AI stories today
Anthropic's IPO roadshow — Your main agent vendor is about to answer to public markets, with quarterly pressure to match. Anthropic
ChatGPT Business Agents — OpenAI is letting verified businesses run their own labelled agents inside ChatGPT — US e-commerce and online services first — plus a Business Hub showing how often the brand gets mentioned. Your support and commerce surfaces may soon need to speak to brand agents, not just websites. The Verge
Claude Fable 5.2 — Anthropic reports its new model scored 64.9% on the RoboDojo robot-control benchmark, more than double the 29.8% it measured for GPT-6 Astra — but Anthropic ran the test itself and the number isn't on the public leaderboard. Treat vendor-run scores as unverified until third parties check them. Humanoids Daily
One takeaway
Audit the boring agent jobs — and ask where your keys live. Today's stories point one way: the costly part of agentic engineering has moved from writing code to checking it, and the savings are hiding in dull, high-volume jobs nobody examines. This week, pick one classification or triage task your team runs on a frontier model and trial a cheaper, verifiable replacement side-by-side on real traffic — and ask security where your team's AI chat logs and provider keys actually live, because two of today's items say that question is overdue. Strands Agents (AWS)