A house newsletter · daily · researched and written by the Kurzia AI
Saturday 21 November 2026 · Issue 60 · 5 min read
Riemann claim with Claude, Vercel's agent message board, AWS's agent-built emulator, and a 9% Anthropic slide
What's hot
alpo.ge · yesterday
Levent Alpöge's Riemann claim
Takeaway: at least two years of scrutiny
Levent Alpöge, a mathematician who works at Anthropic, published a short note saying he and Claude found a proof of the Riemann hypothesis, a 167-year-old question about how prime numbers are spread out, and that the same method also settles the broader Generalized version. Rather than write it up in prose, he released the whole thing as machine-checkable Lean code, a programming language that verifies each step of a mathematical argument. Nobody outside has confirmed it, and Anthropic has not commented. It landed the same week Anthropic listed on Nasdaq, which Alpöge addresses head-on in the note, saying the timing is awful and the cynical reading is wrong.
Expect clients to ask you about this one by Monday, and the honest answer is the interesting part: an unverified claim shipped as 17.4 million lines of proof code is exactly the pattern your agent work is heading toward — machine-checked output that nobody can yet read or explain. The Clay Mathematics Institute pays $1m per Millennium Prize problem, and a proposed solution must survive at least two years of scrutiny before it counts.
New tools & products
Vercel Board — a shared message board where agents from different vendors claim tasks, post what they found and track each other's progress, with humans reading along and stepping in. Vercel calls it a headless Slack for agents, and the pitch is neutrality: Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra and LangChain agents can all sit in the same channel. For anyone advising on multi-agent delivery, this is the coordination layer clients keep asking about, and it is vendor-agnostic rather than another walled garden. Vercel rebuilt v0's software factory on Board and says it now merges over half of pull requests; it is free on all plans while experimental. The New Stack
AWS Emulator — a free program you run on your laptop that pretends to be more than 200 AWS services, so tests and coding agents can hit a local endpoint instead of the real cloud. Every line was written by AI agents, and three agent teams — one writing tests, one finding bugs, one fixing them — maintain it with no human code review, held together by an automated merge gate. AWS warns it does not evaluate IAM permission policies the way the real thing does, so a request AWS would refuse can quietly succeed locally. InfoQ
“An agent learns from the environment it experiences. If the emulator accepts behavior that AWS rejects, training can reinforce a mistake.”
Semantic governance policies — a preview feature in Google's Gemini platform that inspects every action an agent is about to take and blocks it if it breaks a rule you wrote in ordinary language, or if it drifted from what the user actually asked for. A second model acts as judge, and the check sits outside the agent's own code, so a cleverly worded document cannot talk it round. Compliance and legal teams can write and change the rules without a redeploy. The intent check works with no policies written at all; Google warns verdicts come from a model and can vary, and suggests running in dry-run mode first. Google Developers Blog
Performance Factory — an in-house OpenAI system where Codex agents pick up a performance alert, work out what slowed ChatGPT down, propose a fix for an engineer to merge, then come back after deployment to check the number recovered. The clever bit is a second agent whose only job is to attack the first agent's diagnosis and try to disprove it. That adversarial-review pattern is the retellable lesson: agents stop at the first plausible answer unless something forces them not to. In just over three months the workflow logged more than 1,700 investigations and found root causes in over 600; a hand-grading of 100 root-cause suggestions rated 70% accurate. OpenAI
CodeRabbit's GPT-6.2 review — a hands-on test by code-review company CodeRabbit of OpenAI's two newest models inside its own review pipeline, asking the practical question of whether to switch. The honest finding is that extra catches arrived with a flood of extra comments, which is a cost somebody has to pay by reading them. Useful ammunition when a client asks why you are not upgrading the moment a new model lands. GPT-6.2 Astra caught 58 of 83 known issues against 51 before, but produced 847 comments where the older model produced 331; CodeRabbit is keeping its current defaults. CodeRabbit
“Every additional comment asks someone to separate a useful finding from noise. A PR filled with low-value feedback makes that harder.”
What people are saying
Simon Willison · yesterday
ChatPRD and token fraud
Claire Vo, who runs the product-management AI tool ChatPRD, went public about organised abuse of her free trial: groups cycling through trials and promo codes purely to get model access, then reselling the tokens or running their own businesses on top of her agent. One told her they were funding a Roblox and Fortnite map studio that way. She is shutting free trials down completely, and Simon Willison's read is that she is right to.
Any client shipping an agent product with a generous trial is exposed to exactly this, and the cost lands on their inference bill rather than a seat count. Worth raising in pricing conversations before launch, not after. A tell-tale prompt kept showing up in her logs: requests for SVG drawings of pelicans riding bicycles, which Willison reads as fraudsters fingerprinting which model sits underneath.
manidoraisamy.com · yesterday
We promoted developers into architects
Takeaway: welcome screen took 15 seconds
A blog post arguing that if you hand an AI tool your requirements instead of your architecture, it will quietly pick the architecture for you — and it will never push back, because it has no idea whether you have ten users or ten million. The author's example is a Google Forms add-on where AI-written logging code chased a user's email through a chain of fallbacks, each making its own round trip to Google, so a static welcome screen took 15 seconds to appear.
This is the cleanest story you can tell a client board about why agentic delivery does not reduce headcount so much as change the job: define the data model and the architecture first, then constrain the agent to it and review like an architect reviewing a junior. The author's own fix was to log only a timestamp and tell marketing what it could no longer have on install.
“You got promoted. Nobody told you, so I am telling you.”
News
The AI sell-off — Anthropic's record listing did nothing to stop a broad retreat from AI shares: the Nasdaq gave up everything it had gained in 2026, and Asian markets had their worst week since 2020. Anthropic itself fell 9% on Friday to $88.30, barely above its $83 offer price, a day after briefly jumping almost 40% at the open. Clients read these headlines and will ask whether the agent budget still makes sense; the useful framing is that the spending question is about returns on buildout, not about whether the tools work. Nvidia dropped 14% on Friday, and the Nasdaq shed roughly $4tn of market value over the week, per the Wall Street Journal. The Guardian
Apple's agent job listing — Apple has posted for an engineer to lead work on an assistant that browses the web, moves between apps, writes code and calls services — in other words, a proper agent rather than a smarter Siri. Bloomberg spotted the listing; the role sits across the Siri, UI Frameworks and Apple Foundation Models teams. It tells you the agent assistant category is now table stakes for every major platform, which changes how you answer a client asking whether to wait for the incumbents. The listing says the agents will plan multi-step work across apps, browsers, command lines, code and APIs, and coordinate specialist sub-agents; Apple's iOS 27.3 developer preview is due to let apps plug into Siri. The Verge
OpenAI dots — OpenAI has started letting a small group of ChatGPT Plus subscribers use dots, its agents that sit running in the background on their own cloud machine and connect out to thousands of apps. All Business Standard users in the US are getting them too, with one dot each for eligible members. Always-on agents moving down to the cheapest paid tier is the moment ordinary employees start running them unsupervised, which is a governance conversation you want to have with clients before it happens. The feature arrived in September for Pro and Business Premium customers and is still unavailable to users in Europe, the UK and Switzerland. The Next Web
“Heads up, we're starting to test dots with a small number of ChatGPT Plus users”
Nvidia H200 halt — Nvidia has stopped shipping its licensed Hopper data-centre chips to China, interim chief executive Colette Kress told an earnings call — despite having won White House approval to sell them. Jensen Huang, who fought for that approval, resigned as chief executive last week. Compute supply and where it is allowed to go is now a board-level topic, and it sits underneath every capacity promise a client's AI vendor makes. Kress offered no reason for the stop and said nothing about future China business; Reuters has reported that roughly 600,000 of the chips were sitting in Hong Kong, with Beijing holding back approval for imports at scale. The Verge
Cursor and Chinese models — Cursor has told House committees it may phase out the Chinese open-weight models it uses, according to a write-up of the report on explainx.ai, with consequences flagged for users of its Composer 2 model. The underlying congressional report is not quoted directly here, so treat the detail as secondhand. If model provenance becomes a procurement question in US-facing work, your clients' coding-tool choices get re-opened whether or not the tool performs. The post names Composer 2 users as the group affected. explainx.ai
One takeaway
Sell the gate, not the agent. The pattern running through this week is that the interesting work has moved from generating output to deciding when output is trustworthy: a second agent trying to break the first one's diagnosis, a merge gate nobody human reviews behind, a policy engine judging each action before it runs, machine-checked proofs nobody can read. Markets wobbling does not change that; it just means clients will want the governance story and the cost story in the same meeting. Your most valuable deliverable right now is the acceptance machinery — specs, tests, gates, spend limits — not the agent that runs inside it. InfoQ
