Ephemeris · Issue 114 Saturday · 15 Aug 2026 · Zürich

Ephemeris. Nineteen feeds read, four printed.

Filled circles are the sources that made today's issue. Hollow ones we read and set down.

  • The Batch
  • The Zvi
  • Import AI
  • Jeff Bullas
  • OpenAI
  • Anthropic Eng
  • Anthropic News
  • GitHub
  • Sentry
  • Vercel
  • PostHog
  • Cloudflare
  • Fly.io
  • Next.js
  • seeallochnaya
  • TochkiNadAI
  • denissexy
  • rvnikita_blog
  • ProductsAndStartups
Ephemeris · Issue 114 01 / 05 · Anthropic
Provenance · Model behaviour

Claude signs every sentence, and the translation carries it.

Anthropic turned on SynthID-Text across Claude's output and published how it works. The words are not altered. What changes is the source of randomness that picks among them: a cryptographic key plus the preceding context decides the low-stakes choices, leaving a statistical fingerprint in the sequence. Ask Claude to translate a watermarked passage and the mark survives, because Claude chose every word in the new language too.

Holds up against
Light editing
Also survives
Translation by Claude
Method
SynthID-Text, DeepMind, Nature 2024
Erased by
A rewrite that replaces every word
Weakest on
Code, short samples, constrained facts
Detection API
Announced, not yet shipped
Ephemeris · Issue 114 02 / 05 · Cloudflare
Security · MCP

One header gives away every MCP call crossing your network.

Cloudflare Gateway now classifies MCP traffic by reading the MCP-Protocol-Version header on every TLS-inspected request. A new selector puts that in a policy, and a separate traffic-source selector tells an approved organisational portal apart from the server an engineer wired up on their own laptop. Cloudflare is honest about the limit: the header proves MCP, its absence proves nothing. Per-tool control is still on the roadmap.

Gateway policyHTTP · TLS inspection on
# everything speaking MCP on the network
experimental.is_mcp == true

# ...that did not come through an approved portal
and mcp_portal not in {approved}

→ action: block / log / isolate
Ephemeris · Issue 114 03 / 05 · OpenAI
Cost · Agent architecture

Send the extraction to the cheap model. Keep the frontier for the ambiguous part.

OpenAI's builder guide for GPT-5.6 is a routing argument aimed at startups watching their inference bill. Split an agent by task difficulty rather than running one frontier model end to end: classification and extraction go to a smaller tier, planning and genuinely ambiguous requests get the expensive one. The Responses API side adds subagent primitives, so a primary agent can fan parallelisable work into separate workstreams and synthesise what returns.

Step in the agent loopWhere it should run
Classify the incoming requestSmaller tier
Extract fields from a documentSmaller tier
Plan a multi-step taskFrontier
Resolve an ambiguous instructionFrontier
Parallelisable sub-tasksSubagents, then synthesise
Ephemeris · Issue 114 04 / 05 · Cloudflare
Access control · Workers

Your internal app stays public until somebody remembers to lock it.

Cloudflare will now attach an Access policy to the Worker itself instead of to each hostname you remember to configure. One policy covers custom domains, routes, workers.dev subdomains and preview URLs together. Set it at the account level and every Worker in the account is private by default, which is the setting that matters when the person shipping an internal dashboard in an afternoon has never opened the Zero Trust console. Local dev reads the same policy from wrangler.jsonc.

wrangler.jsonc
{
  "access": {
    "dev": {
      "aud": "my-app",
      "identity": { "email": "admin@company.com" }
    }
  }
}
Ephemeris · Issue 114 05 / 05 · The Batch
Craft · Career

Four skills, read out of ten thousand job postings.

Andrew Ng published an AI Engineering Skills Map assembled from more than 10,000 job postings and interviews with practitioners. His argument is that these four belong to every developer rather than to a job title — the same way working with the cloud stopped being a specialism and became table stakes.

  1. IBuild and deployLLM behaviour, context engineering, RAG, agentic workflows, and the statistics to measure and govern what you shipped.
  2. IISoftware fundamentalsReading the tradeoff between cost, scale, reliability and security well enough to choose an architecture.
  3. IIIWorking with coding agentsManaging their context, planning the execution, steering them toward software that survives review.
  4. IVShaping the buildMoving out of pure implementation into product judgement, business sense, and pushing the thing forward.
End of issue 114 Back to top ↑

That's today.

Issue 114, Saturday 15 August 2026. Five picks: a watermark, a header, a routing table, an access policy, and a skills map.

Chosen against the standing rubric — AI tools worth adopting this week, creative and generative software, dev tools and agentic coding, privacy and security where the signal is high, research with a practical kernel, and anything a senior engineer or founder could act on by Monday. Today's picks came from Anthropic, Cloudflare, OpenAI and The Batch.

The Batch · deeplearning.ai Don't Worry About the Vase Import AI · Jack Clark Jeff Bullas · Jabs OpenAI News Anthropic Engineering Anthropic News The GitHub Blog · Engineering Sentry Engineering Vercel Blog PostHog Blog The Cloudflare Blog Fly.io Blog Next.js Blog @seeallochnaya @TochkiNadAI @denissexy @rvnikita_blog @ProductsAndStartups