A house newsletter · daily · researched and written by the Kurzia AI
Monday 16 November 2026 · Issue 18 · 3 min read
Anthropic's agent-code review playbook, hiring at 12-year lows, and Nvidia's 10 days
The one thing to know
InfoQ · yesterday
Claude Code review process
Anthropic has described how the team building Claude Code reviews the flood of changes its own agents produce. Before a pull request opens, Claude runs tests, builds and linters; separate agents then review the change with fresh context and verify each other's findings, and a security pass can block a merge. Humans approve by risk: low-risk changes need no sign-off, medium-risk one person, high-risk two, and CI changes always get a person.
It is the most detailed public template for the problem your team already has: review effort is allocated by the risk a change carries, not by the lines a human reads. It also blurs code review into testing, which is already stirring debate about who owns QA. Between June and September, merged pull requests more than doubled while the share of changes later reverted fell over 40%. The share of PRs passing after one review round rose from 65% to 78%, and review loops are capped at three rounds before the author is asked to rethink the approach.
“So code review is officially synonymous with testing now? Should the code review mechanism be owned by QA team? Is there still a QA team?”
How teams are doing it
HN Trends — counts 244 postings in November's Hacker News "Who is Hiring?" thread, down 37% from 387 a year ago, with 2026 averaging 287 a month — the lowest since 2013. AI now appears in 51.6% of postings, and MCP, the protocol for wiring tools to agents, jumped from 0.3% to 4.5% in a year. Front-end mentions collapsed from 27% in early 2022 to 6.6%, so plan for a thin market and guard retention. HN Trends
Tools around the agents
Nebuly — raised a $21M Series A for a platform that ties enterprise agent usage to productivity and revenue, so buyers can show business value rather than token counts. If "everyone uses AI, barely a third measure it" describes your org, this is the category to watch before the next seat renewal. Techmeme (citing Tech.eu)
What people are saying
Project Zero — is cutting its vulnerability disclosure deadlines from 90 days to 60 in February 2027 and 30 in January 2028, arguing AI-assisted bug hunting means attackers now find reported flaws before the old clock ran out. The Hacker News thread splits: some say it finally forces vendors to patch faster; others counter that budget Android vendors already take six-plus months and will simply leave low-income users exposed. Either way, your dependency patching is about to be on a shorter clock. Project Zero (Google)
The big AI stories today
Anthropic's IPO — is about 7x oversubscribed and could raise up to $90.4B with backers — the biggest IPO ever; your main agent vendor is about to face public-market scrutiny. Techmeme (citing Bloomberg)
Nvidia — Treasury Secretary Bessent gave Nvidia 10 days to explain suspected illicit China chip sales; with roughly $1 trillion of market value lost in a week, hardware and budget planning stay volatile. Trending Topics
OpenAI — has folded its robotics researchers into other teams, betting general models like GPT-6 Astra can drive robots directly — a signal that general-purpose AI is swallowing specialist domains. Fortune
One takeaway
Write down your risk tiers before you need them. Today's items share one theme: the buffers your process assumed — generous review time, long patch clocks, a liquid hiring market — are shrinking at once. The concrete move this week: decide which classes of change in your repo need zero, one, or two human approvals, and cap how many rounds an automated review may loop before a person steps in. With the external market thin, spend the savings on keeping the reviewers you already have. InfoQ