A house newsletter · daily · researched and written by the Kurzia AI
Saturday 21 November 2026 · Issue 23 · 4 min read
AI spend under question after the sell-off; Cursor's China-model problem; agents get a message board
The one thing to know
The Guardian · yesterday
The AI sell-off just became a budget question
Takeaway: $88.30
On Friday the Nasdaq erased everything it had gained in 2026, the S&P 500 closed more than 10% below its record, and Anthropic — two days after the largest IPO in history — fell 9% to $88.30, barely above its $83 offer price. Chipmakers took the worst of it and Asian markets had their worst week since 2020. The through-line is investors openly questioning whether the billions poured into the AI buildout will ever pay off.
For your team this arrives as budget scrutiny: expect finance to read agent seats and token spend with new skepticism, and expect harder funding conditions for smaller vendors you depend on. Nothing about the tools themselves changed — the question is who can show their spend pays. The Nasdaq lost roughly $4tn in market value this week; Anthropic raised $81.34bn at a $1.414tn valuation in its debut.
How teams are doing it
Martin Fowler's Fragments — rounds up an experiment by Marcus Borg: three LLMs reviewed and fixed 465 pull requests across 246 open-source repositories, and a typical unhealthy PR cost 22 to 35% more to push through the review-and-fix pipeline than a healthy one. Refactoring stops being hygiene and becomes a line in your agent bill — the messy corners of your codebase are exactly where review loops will burn tokens. Complex methods and complex conditionals were the smells that raised cost most, per CodeScene's health scores. martinfowler.com
AI picks your architecture if you don't hand it one. A developer found a colleague's AI-written logging code was making several Google API calls on every install, stretching the addon's welcome screen to 15 seconds. The fix an architect would choose — log a timestamp, show the screen first, tell marketing no — is one the agent never proposed, because it executes requirements without pushing back. The author's rule: specify the architecture and data model first, then let the AI code inside those constraints. Worth adding to your review checklist for agent-written code: check design choices — hidden API-call counts, duplicated data models — not just correctness. The author's summary of the role shift: the coding job is going away, but the architect's job is arriving, unannounced. manidoraisamy.com
Agents and plans
Cursor — told US lawmakers it is considering phasing out Chinese open-weight AI models over what it calls geopolitical risk, CNBC reports — notable because its Composer 2 model was built on Kimi K2.5, a Chinese open-weight model. Your agent vendor's model supply is now a policy question: the model underneath a tool you standardised on could be swapped or restricted without your say. Ask vendors where each model comes from before you renew. The disclosure came in statements to lawmakers rather than a product announcement, so treat the phase-out as under consideration, not decided. CNBC via Techmeme
ChatPRD — founder Claire Vo says organised groups abused its free trial at scale — looping trials and promo codes to run commercial operations, including a Roblox and Fortnite map studio — and she is winding down free trials entirely. Expect free evaluation of AI tools to get scarcer: when you vet the next agent product, budget for a paid pilot instead of counting on an open-ended trial. One recurring probe prompt was SVGs of pelicans riding bicycles, used to fingerprint which model sat underneath the service. simonwillison.net
Agent Substrate and gVisor — were donated to the CNCF, the vendor-neutral home of Kubernetes, alongside Google's OpenRL. gVisor isolates untrusted agent-written code from the host kernel; Agent Substrate parks idle agents cheaply and restores them fast instead of keeping a container running for each one. If you are building an internal agent platform rather than buying seats, this is the layer becoming standard, and open governance cuts lock-in risk. Google estimates agents sit idle 90-95% of the time while holding state, and says a paused agent resumes in under 500 milliseconds. InfoQ
Tools around the agents
Vercel Board — is an experimental shared board where agents from different tools — Claude Code, Codex, LangChain and others — claim tasks, post findings and request human approval. Vercel rebuilt its v0 software factory around it and says over 50% of PRs now merge. A readable pattern for supervising many agents at once instead of five chat windows — with a caveat the docs are blunt about: agents may read any message as an instruction, so keep keys and tokens off the board. It is free during the experimental period and available on all Vercel plans. The New Stack
Groundcover Investigations — is a shared incident-response space where an investigator agent tests hypotheses against production telemetry and a separate judge agent must approve every finding before it can close; on-call engineers watch each hypothesis live and can steer or override at any point. The doer-judge split is a pattern you can borrow for your own review rules, and it addresses the real on-call risk of agents acting beyond their scope. Coding agents like Claude Code and Codex can join an investigation through an MCP server; Salesforce says organisations already run an average of 12 agents. The New Stack
Machine-checked proofs aim at AI-written code. A survey of startups — Harmonic, Theorem, Axiom Math and new entrant &Truth — building tools that prove code correct with Lean, a language that machine-checks every step of a mathematical argument, pitched as the answer to low-quality AI-generated code. When a coding agent says a change is verified, it means it ran tests; these tools offer a stronger guarantee, and if the cost keeps falling it changes what you can demand of agent-written PRs. Anthropic reported Opus 5.5 solved 50% of 36 small verification tasks using about 40% fewer tokens than its predecessor. Latent Space
What people are saying
Reviewing AI's code won't keep your reviewers sharp. A widely shared essay argues the opposite of emerging practice: engineers who stop writing code get worse at reviewing it, because skill needs practice like any craft, and passively reading agent output doesn't count as practice. It pushes back on the assumption that humans can simply become reviewers of agent-written code. If there is truth in it, teams need to keep real hands-on problem-solving in the job, or review quality decays quietly over a year. The author likens it to a guitarist who stops practicing and only directs other players, then judges the results by ear. cekrem.github.io
The big AI stories today
Riemann claim — Anthropic mathematician Levent Alpöge says he proved the Riemann Hypothesis with Claude, released as machine-checkable Lean code — the highest-profile claim yet of AI as research collaborator, and awaiting verification. alpo.ge
Signal — now blocks screenshots on iOS 27 and Android by default to keep OS-level assistants like Siri and Gemini from reading chats — expect staff privacy questions. Signal
Cognition — Devin-maker Cognition is reportedly in talks to raise over $2B at a $75B valuation, two months after raising $2B at $48B. Bloomberg via Techmeme
One takeaway
Pull your agent spend numbers before finance does. The sell-off doesn't change what your agents do this week, but it changes how finance reads your line items — expect questions about seats, tokens and what they bought. Before that conversation arrives, put cost per task or PR in front of whoever holds the budget, and pair it with the code-health finding from today's practice section: cleaning up the messiest parts of the codebase is now a measurable cost cut, not just hygiene. martinfowler.com