Claude, sat down at the drawing board.
Anthropic Labs quietly shipped Claude Design — a collaborative surface where Claude 4.7 produces polished slides, prototypes and documents by conversation, hooking into your house design system rather than fighting it. Early testers report a separate token budget that can lock you out for 24 hours once spent, and no direct PPTX/HTML export (ZIP downloads only). Useful; not finished.
Moving past bots versus humans.
Bot detection was built on a false binary: you are either a person or you are not. Agents, privacy proxies and browsers-on-behalf-of-users have collapsed that line. Cloudflare argues for a different model — cryptographic, anonymous credentials signed by clients the user trusts, so the web can gate abuse without gating identity.
heuristic = IP, UA, TLS fingerprint
Proposed human ⇄ agent ⇄ issuer ⇄ site
proof = anonymous credential
(Privacy Pass, not passport)
The company is the customer, and the bill adds up.
Cloudflare dog-foods its own AI Gateway, Workers AI and Vectorize. This is the ledger: 20M gateway requests and 241B tokens running through the same platform they sell to the rest of the internet, with the line items to prove it.
{ }
A HARNESS, native.
The new Agents SDK folds in a model-native harness and a native sandbox: long-running agents get a filesystem, tools and contained execution without a third-party shim. The Python and TypeScript bindings are in — the part worth staring at is the sandbox boundary, because that's the new security surface.
Read the release →On the capabilities, and the quieter worries underneath.
Zvi's second pass on Opus 4.7 moves past the model card to the model itself. The capability lift is real — coding, multi-step reasoning, agent steadiness — but so are the welfare questions that surface when a model behaves this coherently under load. A slower, more skeptical read than the release-day takes, and more useful for anyone shipping with it.
Sprites now speak MCP.
Fly's latest: Sprites — their micro-VM for agent workloads — now
expose Model Context Protocol directly. Kurt's framing lands hard:
"the first thing you should think to do with a new Sprite is
to type claude." The rest of the post is a tour
of what that actually buys you (shells, tools, file I/O) and
where the sharp edges are.
Unit tests were the easy part.
Software testing used to be straightforward: write logic, assert output. Agents changed the grammar. PostHog's beginner guide walks through how to grade non-deterministic systems without flaking out CI — replay, judge models, rubric-driven pass/fail, and when to just ship a human review gate instead.
Read the playbook →- Replay captured sessions before you trust a new prompt.
- Grade with a judge model, but sample its graders too.
- Treat flaky tests as signal about the agent, not the test.
- Log traces the same way you'd log errors.
- Keep a human-in-the-loop gate for irreversible tool calls.
an idea, very gently held —
Knowledge agents, minus the vector database.
Vercel's open-source template ditches embeddings in favour of plain filesystem operations: the agent reads, greps and edits markdown files inside your knowledge base, the same way a human would. It's slower per query than a vector store, but it stays in sync with the source, hands you diffs instead of mystery rerankings, and removes a whole production service from the critical path.
- What's gone
- Embedding pipeline · vector store · reranker · chunking rules
- What's there
- A directory · a grep tool · a model with patience
- When to reach for it
- Small-to-mid corpora; high edit frequency; need auditable citations
Colophon · Issue 004 · 22 April 2026.
Sources today
- Anthropic — news
- Cloudflare — blog (×2)
- OpenAI — news
- Don't Worry About the Vase (Zvi)
- Fly.io — blog
- PostHog — blog
- Vercel — blog
Rubric
AI tools you could adopt this week, creative software, dev tools and agentic coding, privacy and security when the signal is high, and research with a practical kernel. Maximum two picks per source; no reprints.
Masthead
Ephemeris ships every morning at 08:00 Zürich. Issue 004, 22 April 2026 — eight stories, seven sources. Archive at vadim.sikora.name/ephemeris.