Ephemeris.
Ten stories from six corners of the internet, picked so you can spend the rest of the day making things instead of reading about them.
A model that does what you say.
Katie Parrott spent a week putting Anthropic's new Opus 4.7 through Every's usual gauntlet — writing, synthesis, hairy code. Her verdict: it's more literal, more precise, and the best model they've tested on well-specified tasks. The catch is that "well-specified" does more work than it used to. Bring your spec; leave your vibes at the door.
Read on Every →The folder is the agent.
Kieran Klaassen's argument is almost absurdly simple: stop juggling system prompts across tabs. Give each model its own folder, let the folder define the context, and the thing you were calling "an agent" turns out to be a directory you already know how to manage.
The workflow is the folder. The folder is the agent.Read the playbook →
4★
Four agents run a 25-person company.
Parrott lays out the four custom agents Every now runs on: prioritization, meeting notes, OKR planning, and a fourth for editorial review. The post ships with reusable prompts, which is the point — this is a recipe, not a brag.
- 25
- Humans on staff
- 4
- Custom agents
- ∞
- Prompts shipped
Marie Kondo your Mac.
Yash Poojary relaunches Sparkle — a tiny utility that watches your Mac and quietly files things where they belong, so the Downloads folder stops being a black hole. No AI hype in the copy, which is refreshing. It just moves files.
Pick it up if you have more than 2,000 items in ~/Downloads and a vague suspicion that tidying is a solved problem that nobody solved.
╭─────────────────────╮ │ ✨ sparkle.app ✨ │ │ watching... │ │ ▓▓▓▓▓▓▓░░ 73% │ ╰─────────────────────╯
The week agentic coding moved a gear.
Zvi's seventh installment on agentic coding tracks the arrival of "Auto Mode" across Claude Code and Codex — where the agent stops asking and starts running. He argues the productivity curve is now real, but so is the skill gap between people who set up guardrails and people who do not.
The pattern across every Auto Mode demo is the same: when a developer narrows the blast radius — sandboxes, feature branches, pre-committed tests — agents ship more in an hour than most engineers ship in a day.
When the blast radius is the whole repo, agents still ship more in an hour. The work just isn't always work you wanted.
Zvi's take: the tooling is finally good enough that resisting it is a political, not a technical, stance.
AI agents can now reverse-engineer you.
Jack Clark's headline result this week: Claude Opus 4.6 reimplemented a 16,000-line bioinformatics toolkit from the binary alone — weeks of human work, compressed into an overnight run. He pairs the benchmark with ten scenarios for "gradual disempowerment" that are less sci-fi than you'd hope.
Scaling laws for cyberwar.
Frontier models now clear 50% on cyber-offense tasks that take expert humans three hours or more. Clark walks through the curve and what it means for defenders (who are the same humans, on the same clock). MIT research in the same issue sketches the slower, broader wave of text-task automation landing by 2029.
Codex leaks, Gemini sings, LLMs learn on the fly.
- Claude Code's source code surfaces in a leak — and the internals are more boring than people hoped.
- OpenAI quietly exits consumer video generation, handing the field to Runway and friends.
- Gemini ships a music-generation module you can actually hum along to.
- New papers show LLMs can meaningfully learn at inference time, not just retrieve.
There is a superapp hiding inside Codex.
The Rundown notes that OpenAI has quietly shoved browser control, shell, and agents into Codex — crossing from IDE plug-in into operating layer. Paired item in the same issue: Ollama now handles local-LLM deployment on laptops, so there's a matching off-ramp for anyone who wants the power without the SaaS.
↳ opening browser...
↳ comparing 4 destinations...
↳ drafting itinerary...
✓ booked. receipt in inbox.
$ _
Stop using just one AI.
Bullas's argument: one chatbot is a Swiss army knife — fine until you need to actually cut something. Ten specialists beat one generalist for most real jobs. Here's the spread.
That's all for today.
Ten picks from Every, The Batch, The Rundown, Don't Worry About the Vase, Import AI, and Jeff Bullas. Back tomorrow at 08:00 Zürich.
Sources
every.to
deeplearning.ai/the-batch
therundown.ai
thezvi.substack.com
jack-clark.net
jeffbullas.com/jabs
Rubric
AI tools · creative software · dev tools · privacy · science · the practical. If we can't imagine you using it tomorrow, it doesn't run.
Colophon
Set in Fraunces & Inter (Google Fonts), with JetBrains Mono on the terminal pages. Hand-laid in HTML/CSS. Issue 001 assembled 08:00 CET, Sun 19 Apr 2026.