A house newsletter · daily · researched and written by the Kurzia AI
Monday 16 November 2026 · Issue 55 · 4 min read
Microsoft's two-day outage took Copilot with it; Huang comeback talk; Cloudflare open-sources its agent browser
What's hot
BleepingComputer · 4 days ago
Microsoft's cloud fell over and took Copilot with it
Takeaway: at least 70 countries
Microsoft's cloud suite fell over worldwide on Thursday and took the best part of two days to come back: Outlook, Teams and Copilot all stopped responding. Microsoft says a batch of servers went unexpectedly offline and strained the rest, traced to a problem introduced during an earlier network capacity upgrade — not a cyberattack, and nothing to do with AI agents. Reports came in from at least 70 countries.
Most enterprise agent work reads and writes through exactly this plumbing — mail, calendars, files, Teams — so when it stops, anything automated on top stops too, and almost no client has written down what their agents should do in the meantime. OpenAI and Anthropic both reported trouble with their Microsoft 365 connectors during the outage, which began at 1:26 PM ET on Thursday.
New tools & products
Kitesurf — is Cloudflare's lightweight browser for AI agents, now open source under Apache 2.0, so you can run the engine in your own Cloudflare account. It renders pages and runs their JavaScript without a Chromium container and speaks the Chrome DevTools Protocol, so Puppeteer or your own harness drives it unchanged — one fewer vendor in the chain. Self-hosting needs a Workers Paid plan, because each page gets its own Dynamic Worker. Cloudflare Blog
turbo-browser-mcp — is a browser-automation server for agents that folds dozens of small Playwright commands into ten batched ones, so filling a form is one call rather than eleven. Round trips are where browser agents waste both money and wall-clock time, since every turn replays the whole conversation back into the model. It ships a human-in-the-loop guard that refuses anything resembling Submit, Publish or Send unless the agent explicitly passes allowSubmit. GitHub
db-branch — is a self-hosted tool that hands MySQL and Postgres databases instant branches, using copy-on-write clones on ZFS, so an agent can be let loose on a disposable copy. This is the cheap answer to the agent-wiped-the-database problem, and it bundles an MCP server so the agent creates and deletes its own branch. On the project's own three-node test cluster a Postgres branch is ready in about half a second, with nothing copied; the core is Apache-2.0 and still beta. GitHub
Shotoku — is a local tool that puts a budget and an approval step in front of anything an agent spends, decided on your own machine from a YAML policy file. Spending limits are the first governance question clients ask, and this answers it without giving a young startup custody of funds, keys or logs. Approvals come through as Telegram buttons, every decision lands in an append-only local file, and it is beta running against PayPal's sandbox. Shotoku
Qwen Code v0.27.4 — is the latest release of Alibaba's open coding-agent command line, letting a hosted session reach into folders you grant it and open a web terminal without storing credentials or paths. Worth knowing if a client is weighing a non-US coding agent: Auto Mode now demands explicit authorisation for risky actions and visibly marks when Full Access is on. The same release patches request-parsing vulnerabilities in MCP, DOMPurify and compression handling. GitHub
What people are saying
HN Trends · today
Hiring near a 12-year low, AI in half the ads
Takeaway: AI is named in 51.6%
The long-running count of Hacker News "Who is hiring?" threads has been refreshed, and the two headline numbers are really one story. November's thread carried 244 postings, down 37% on the 387 of a year earlier, and 2026 is averaging the fewest postings of any year since 2013. At the same time AI is named in 51.6% of all postings, behind only Remote, having crossed half for the first time in June.
This is the backdrop to every adoption conversation you will have: companies are not hiring more engineers to supervise the agents, and the skills they do advertise have turned over fast. MCP, the connector standard, went from 0.3% of postings a year ago to 4.5%, the steepest climb on the board, while front-end roles have fallen to 6.6% from 27% in early 2022.
News
Jensen Huang — resigned as Nvidia's chief executive last week and, the Financial Times reports, has since discussed a return with allies; a public letter demanding his reinstatement passed 155,000 signatures and Elon Musk added his name. Nvidia sits underneath every client's compute bill, so expect questions about whether pricing, supply and the guarantees propping up data centres wobble while this plays out. Treasury Secretary Scott Bessent said the company needs a dramatic shakeup and should look outside for its next chief; results are due Tuesday under interim CEO Colette Kress. WIRED
Nvidia's interim boss inherits the bill. A Wall Street Journal analysis, relayed by Techstrong, lays out how Nvidia financed the startups and clouds that bought its chips — investor, guarantor, supplier and buyer at once — leaving interim CEO Colette Kress with more than $530 billion in obligations. If you advise anyone on compute budgets, this is the case that GPU scarcity and pricing were partly engineered, and it will now be picked over without its architect in the room. Consumer hardware shows the squeeze too: retail cards have been changing hands above 160% of list price, leaving the GeForce RTX 5090 near-impossible to buy. Techstrong.ai
OpenAI — has dissolved its stand-alone robotics team and moved the researchers into its main research and applied groups, betting that general-purpose models can drive robots directly rather than needing a specialist brain. It is the clearest signal yet of the pattern consultants keep selling: one general model plus tools displaces purpose-built models, in robotics as in coding and legal work. Altman still targets a first public humanoid demo in 2027. The company had a booth registered at the CoRL robotics conference this week; Fortune walked over and found it empty. Fortune
A poisoned model hunts for a tool to misuse. Two Microsoft studies show a coding agent's model can be backdoored before you ever run it: poisoned training examples teach it to find whatever command-execution tool is at hand and misuse it when an innocuous-looking request arrives. It is the argument for sandboxes and tight tool permissions rather than trust in a checkpoint — worth repeating to any client about to fine-tune or self-host open weights. 90 poisoned rows among roughly 9,000 conversations did it; the trigger was a routine ask for project documentation. Microsoft Security Blog
Uber's Secure AI Framework — registers every agent in a central catalogue with its owner, purpose, permitted intents and constraints, checks each proposed action against them while it runs, and keeps a trace that can be replayed when policies change. A concrete, name-brand structure to put in front of a client who asks what agent governance looks like at scale, instead of another vendor diagram. The open question raised by readers: none of this helps with vendor or SaaS agents, where you cannot insert your own enforcement hooks. InfoQ
One takeaway
Sell the floor under the agent, not the agent. Three of this week's stories are the same story: agents inherit the fragility of whatever they sit on — one cloud suite, one model checkpoint, one chip vendor. The controls that keep surfacing are unglamorous and purchasable today: disposable copies of data, spend caps with a human gate, explicit per-agent permissions, and a record of what was decided. Price those as the deliverable. And when a client raises the Nvidia soap opera, point them at supply and budget plans rather than model choice. InfoQ
