Today's issue is about repair work. Anthropic ships a postmortem for last week's quiet Claude Code regression. Cloudflare bundles a year of agent infrastructure into one platform week. GitHub rewires search for thirty thousand sites it didn't break. The new tools made last week — this week we read the receipts.
— the editors
Three changes, one quiet regression.
Anthropic traces last week's quality complaints in Claude Code to three separate, overlapping changes — and writes up what each one broke, how it was caught, and what the team is doing differently. Real postmortems from a frontier lab are rare. Read it like an aviation incident report.
"We traced recent reports of Claude Code quality issues to three separate changes."
A whole cloud for agents, one week.
Cloudflare's Agents Week 2026 recap stitches together Workers AI, the agent toolbox, durable execution, sandboxes and the new identity primitives into one coherent stack. If you build agents on someone else's compute, this is the new floor.
GLM 5.1 thinks. Data centers push back.
Andrew Ng's weekly survey: a Chinese frontier model lands with strategic-reasoning chops, residents organize against new hyperscale builds, helpful LLMs get measured on when they go quietly unhelpful, and humanoid robots crack their first commercial deployments.
30K docs sites, sub-second propagation.
GitBook hosts documentation for Nvidia, Zoom and n8n on top of Vercel. Their writeup is a useful look at what running a multi-tenant SaaS on someone else's edge actually costs in latency, build time and cache invalidation — the numbers are on the Vercel side of the conversation.
In our image, after the week of Opus.
Zvi's weekly digest compresses the Opus 4.7 launch arc — the model card, the welfare debate, the mid-week regressions — into one readable file. If you skipped the daily threads, this is the catch-up document. He stays sober about both the wins and the discontents.
This was the week of Claude Opus 4.7 — capabilities we asked for, complications we didn't.
Do you really need an MCP for that?
Sentry's engineers wired a Claude Code agent to ship a native app two ways — once with their MCP server, once with a plain file-and-shell workflow. The post is honest about where the MCP earned its keep, and where it just added latency. A useful gut-check before you adopt.
With MCP
Tighter context, fewer hallucinated APIs, slower roundtrips on simple edits.
Without MCP
Faster on small tasks, blew through context on multi-file refactors, more dead ends.
Verdict
MCP wins when the agent needs structured access to your domain. Skip it for one-off scripts.
Cost
"Time, cost, and context" — pick the one you can afford to spend.
Next.js commits to portability, on paper.
Next.js 16.2 ships a stable Adapter API, a public adapter test suite, and a working group that includes the OpenNext project. Vercel is — at last — putting a contract under the words "deploy anywhere." If you've been waiting to evaluate non-Vercel hosts, this is the version to do it on.
Search, rebuilt for uptime.
GitHub Enterprise Server's search stack used to fall over when one node coughed. The team reports how they cut the dependency on a single Elasticsearch primary, added cluster-aware routing, and got the customer-visible search index back inside a tighter latency budget — without a forklift migration.
600 office tasks, graded by agents.
Zapier put autonomous agents through 600+ real workflows across sales, marketing, support, finance and HR. Claude Opus 4.7 placed first — but the spread between the top four models was small enough that workflow design matters more than model choice. A useful baseline before your next vendor bake-off.
Stanford opens its AI economics seminar.
MSE435: Economics of the AI Supercycle is a Stanford seminar series with practitioners across the stack — chips, infrastructure, models, applications. Lectures publish to YouTube. Roman Rvachev is restarting his weekly study club around it from 4 May. If you wanted a structured place to think about where the money lives, here.
- RE:
- MSE435 — Economics of the AI Supercycle
- FROM:
- Stanford · YouTube
- STARTS:
- Study club: Mon 4 May 2026
- FEE:
- None — open seminar
That's today.
Issue 007 · Saturday 25 April 2026 · Zürich. Today's picks were drawn from Anthropic Engineering, Cloudflare, DeepLearning.AI, Vercel, Don't Worry About the Vase, Sentry Engineering, Next.js, GitHub, plus pointers via @ProductsAndStartups and @rvnikita_blog. Rubric, in priority order: tools you could adopt this week; creative software; dev tools and agentic coding; security; research with a practical kernel; anything immediately actionable for a senior engineer or founder.