A house newsletter · daily · researched and written by the Kurzia AI
Friday 20 November 2026 · Issue 22 · 3 min read
Anthropic's first trading day, a CEO who dropped PR review, and Gemini 4 Neon lands in Copilot
The one thing to know
Techmeme (Reuters) · yesterday
Anthropic's Nasdaq debut
Takeaway: up 17% at $97.17
Anthropic closed its first day as a public company up 17% at $97.17 on Thursday, after opening at $111.50 — 34% above its $83 offer price.
For teams whose coding agents run on Claude, the practical read is vendor stability: a public Anthropic now has real capital ($107.3 billion including a private placement) and more than $900 billion in compute commitments to meet — but also quarterly pressure to grow revenue, which tends to show up first in seat and usage pricing. Roughly 316 million insider shares can be sold from day one, and the co-founders' Class F stock controls 50.1% of votes, so public investors will have little say.
How teams are doing it
CodeRabbit's agent factory — CodeRabbit CEO Harjot Gill runs 40 engineers as a software factory: agents pushed more than 5,000 pull requests in one month on an AI bill of roughly $250,000 a month. When volume explodes, per-PR review stops scaling, so his engineers review releases, not individual PRs; a review-exempt list lets trusted devs push straight to production, and outages earn strikes that bring review back. A CI vendor quoted $13M a year to handle that load; his agents built a replacement system in a week for half the price. The Pragmatic Engineer Podcast
Adam Tornhill's code-health case — Code-health researcher Adam Tornhill told GOTO Copenhagen that across 1,000+ open-source projects he studied, AI's speed gains were fully erased within two months by rising complexity. It is the argument for a refactoring budget: agents fix only about half the health issues in messy code, but a codebase fed its own health data through an MCP tool reaches a 90–100% fix rate. Healthy code also costs 40–45% fewer tokens; the average codebase scores 5.15 on his health scale, where AI-friendly code starts at 9.5. InfoQ
Agents and plans
Gemini 4 Neon in GitHub Copilot — Google's new mid-tier model — $3 per million input tokens, $15 per million output — is now generally available in GitHub Copilot on Pro+, Max, Business and Enterprise plans, billed at list price under usage-based billing. If you run Copilot for a company, note that new models arrive enabled by default; Business and Enterprise admins have to switch them off through the model policy if they want to control what teams use. Google claims quality close to its bigger Carbon model at a 32% lower cost per task. GitHub Blog
Qwen4-Max — Alibaba's model, at $1.50 per million input and $4.50 per million output tokens, went head-to-head with Claude Opus 5.5 ($4/$20) in three hands-on agent tests at The New Stack, scoring 170 of 183 points against Opus's 179. The cheap option costs time and accuracy instead of cash: it ran roughly three times slower, and the author's advice is to keep the premium model for code and route investigation-style work to the budget one. On the incident-investigation test it actually beat Opus 59–58 for roughly a third of the cost. The New Stack
GPT-6 Luna in Copilot — One of GitHub Copilot's models was degraded for roughly forty minutes on Nov 19 because of a problem at an upstream model provider; GitHub told users to pick another model or switch to its Fusion blend. Every model provider wobbles independently, so a team standardised on one model inherits its outages — this is the drill for naming a fallback. The incident ran from 15:56 to 16:37 UTC and a root-cause analysis is promised. githubstatus.com
Tools around the agents
Devin Bots — Cognition launched Devin Bots, persistent agents that oversee fleets of Devin coding sessions for days or weeks; one bot built Cognition's Android app over 44 days, merging 1,134 pull requests while its supervising engineer sent messages on just 22 of those days. The open question for any buyer is where oversight actually sits: Cognition hasn't said how a bot decides code is merge-ready, or how much engineer time the reviews cost. A mobile game studio running about 20 bots under an agent called Super Jenny went from roughly 50 PRs a week to 1,900. The New Stack
What people are saying
An engineer blasts AI safeguards that blocked a security fix. A widely shared rant from an Estonian engineer describes asking Claude Code for easy security improvements, having it find a real one-line issue, and then getting blocked by Anthropic's safeguards. His complaint: the exemption path takes days, wants a passport or ID, and forfeits data-retention exemptions, while attackers simply use unrestricted models anyway. Worth deciding your security team's approved fallback before they hit the same wall. He ended up finishing the fix with EU-hosted open-weights models (GLM 5.3) instead. ounapuu.ee
The big AI stories today
Apple's agent project — A job listing shows Apple building an autonomous agent that writes code, browses the web, controls apps and coordinates subagents — a fourth giant entering agents. Techmeme (Bloomberg)
Gemini for macOS — Google's Gemini desktop app on macOS can now drive apps, files and the web in the background for US AI Plus, Pro and Ultra users — consumer computer use goes mainstream. Techmeme (TechCrunch)
Nvidia — The SEC is reportedly probing whether Nvidia misled investors, per the WSJ, after Jensen Huang's exit — a governance wobble at the vendor behind most AI compute. TechCrunch
One takeaway
Pick your review unit before volume picks it for you. Between a CEO dropping per-PR review and a researcher showing speed gains eaten by complexity, the through-line is that review and maintenance — not writing code — are now the scarce resources. This week, decide what unit your team actually verifies: a pull request, a release, or a scheduled job's output, and put one automated quality guardrail in front of agent volume.
And don't spend the week reacting to IPO headlines; nothing about your vendor choice has to change because the ticker changed. Techmeme (Reuters)