A house newsletter · daily · researched and written by the Kurzia AI
Saturday 14 November 2026 · Issue 16 · 4 min read
Claude's two outages in a day, Nvidia fallout for your vendor list, and a $3B bet on agent ops
The day's one thing
Claude Status · yesterday
Two Anthropic incidents took agents offline yesterday
Takeaway: more than two hours late
Anthropic's platform failed twice in one day: a morning authentication outage locked users out of claude.ai and Claude Code, and from 13:09 UTC a second wave of elevated errors hit the web app, the API, Claude Code and connectors before being marked resolved at 15:38 UTC. For a team that has agents woven into the working day, that is most of a day gone. The sensible stance is not to leave Anthropic over it — it is to name a fallback and design agent workflows that tolerate the platform being down.
When one vendor sits under your coding agents, its status page is effectively your engineering availability. During recovery, scheduled tasks and routines were running more than two hours late and some runs were skipped outright; connectors, Managed Agents and Security scans stayed degraded.
How organisations are changing
Dario Amodei — says Anthropic turns away deeply technical candidates who don't fit a low-ego, high-trust culture, and spends 40 to 50 percent of his time on talent and culture ahead of its IPO. The payoff, he argues, is a team that stays. When a few senior people carry most of your AI capability, culture-fit hiring is a stability control, not softness. Headcount grew from about 940 at end-2024 to about 5,300 by late September, on roughly 1% regrettable attrition a year; all seven cofounders remain. Business Insider
“We don't want people who are jerks. We don't want people who try to steal the show and don't work well with others.”
Data Intelligence Engineer — a role MotherDuck's Jordan Tigani says is emerging as agents take over writing SQL, dashboards and pipelines: the human curates shared context, runs evals, and stays accountable for the numbers being right. If your data team runs on a ticket queue, expect that queue to empty; the surviving work is verification and context — a vendor's blog, but the direction matches what teams report. His lead engineer says it's the first stretch of his career with no backlog. MotherDuck
“Agents need supervision proportional to their blast radius.”
Vendors, pricing and risk
Microsoft 365 — kept failing for a second day, disrupting governments, hospitals and businesses worldwide, before Microsoft said availability was improving as it worked to restore Outlook and Teams. Office-suite reliability is now agent reliability: whatever assistants your people use live on the same estate, so degraded-mode plans belong in your agent design. Microsoft had not declared full restoration. Gizmodo (via Techmeme)
Fireworks AI — Uber's AI-inference provider — was breached in October, exposing thousands of drivers' licence numbers; Fireworks attributes it to criminals trading stolen tokens for access to large language models. Third-party AI vendors hold your data and your model credentials together; this is the exposure pattern to audit across your own stack. State filings count stolen licences at 2,567 in Massachusetts, 3,430 in South Carolina and 1,873 in Rhode Island; Notion said its data was taken in the same hack. TechCrunch
News the board will ask about
Nvidia — is now a line item in vendor-risk reviews: analysts argue Colette Kress's appointment as interim chief, plus expected harder scrutiny of Nvidia's China dealings, should prompt buyers to reassess how much of their AI infrastructure, suppliers and investment plans rest on one chip company. Concentration in compute is the exposure your board will raise this week; the analysts' answer is measured diversification, not panic exits. Nvidia's revenue more than doubled year over year to $96.2 billion last quarter — the dependence behind the argument. CIO Dive
“treat Nvidia's CEO transition as a prompt to reassess concentration risk across their AI infrastructure, suppliers, and investment plans”
LangSmith — LangChain's platform for running, testing and supervising agents — has a $300 million Series C behind it at a $3 billion valuation, with 7,000 customer organisations; Workday says an agent that once needed five months to ship now goes live in under a day. The same wave funded agent-security startup WhiteMatter's $20 million Series A and revenue-agent vendor Rox's $100 million Series B at $1.3 billion: the layer above the models is where enterprise spend is heading. CapitalG led the round. LangChain
Amazon — is building, per Reuters, a near-universal record of merchandise sold anywhere — including data taken from competitors' storefronts — on the expectation that shopping agents will steer buyers to whichever merchant's product information is richest. The agent-mediated web is arriving from the buying side; worth checking whether your own public product and pricing data is clean enough for a machine to act on. Reuters reports this from an internal document; Amazon has not confirmed it. Reuters (via Techmeme)
The big AI stories today
ChatGPT Wallet — lets OpenAI's agents research and buy items for US ChatGPT Pro subscribers, each checkout approved with Face ID or a fingerprint and card details tokenised — agentic commerce reaching real consumers, while Amazon blocks agents from its own shelves. TechCrunch
Nvidia — held $48 billion of stakes in 229 private companies, including OpenAI and Anthropic, and had offered $165 billion in cloud backstops — a gauge of how much of your AI supply chain one company underwrites. New York Times (via Techmeme)
OpenAI folds its robotics team into research. OpenAI has shut down its standalone robotics team and moved the researchers into research and applied units, on the argument that models like GPT-6 Astra now apply to robots directly — a bet that general models, not robot specialists, carry the field. Fortune (via Techmeme)