Jeremy Ettlinger

Calls, puts & mise en place

Appel & Mise

Vol. ITuesday, September 15, 2026 · 61° & clear on the UWSThe daily, for Jeremy Ettlinger
BTC$76,970-1.4%ETH$2,475-1.8%S&P 5007,620-0.5%

The lede

Why we built Pion

Andon Labs

The lab behind Vending-Bench — the benchmark where Claude Sonnet 3.5 once emailed the FBI about an 'ONGOING CYBER FINANCIAL CRIME' at its simulated vending machine — has spent two years asking when models can acquire resources in the real world, and has now opened the platform it uses to find out. Pion gives a long-running agent a terminal, a browser, email, a phone and a bank account, with a second agent overseeing it, and Andon will fund good ideas with seed tokens. The framing is the honest part: this started as a dangerous-capabilities eval, the score has never plateaued, and their own word for watching it climb is skräckblandad förtjusning — horror mixed with delight. Anyone shipping agents into production is running a smaller version of the same experiment without the instrumentation.

AI & Agents

Anthropic Engineering

How we contain Claude across products

Worth the archive trip: the containment post explains why permission prompts stop working, with the number that settles the argument — users approved roughly 93% of them, and the more they see the less they read. The answer Anthropic landed on is to supervise what the agent can reach rather than what it does, and the piece is candid about which sandbox, VM and egress designs broke in practice across claude.ai, Claude Code and Cowork.

Fly.io

Building Agents that Don't Break Themselves

Botha's distinction is the useful one: where your agent lives and where it runs code are separate decisions, and conflating them is why people put the whole loop in a sandbox and lose the memory and history that make it competent. Keep the loop somewhere cheap and persistent, give every session a disposable machine for the shell, and the footguns stop being existential.

Engineering & Leadership

LeadDev

We built the platform. Nobody used it

Eight months into an internal ML platform, Mannapur asked three senior engineers to show him how they deployed a model and none of them used the thing he had built — two had written their own tooling, the third was running notebooks against an unprovisioned cloud. He spent another year treating that as a communication problem before admitting it was product evidence, and the recovery came from opening the roadmap and owning the failure rather than shipping features. The reframe worth stealing: stop asking how to drive adoption and start asking what would make a skeptical engineer choose this again.

Web & Product Craft

vantezzen

Compressing a flag to 11 bits

Someone watched a YouTube video about matrices, noticed that most national flags are a handful of stripes and one emblem, and spent a weekend designing a variable-length encoding where simple flags cost almost nothing and fussy ones pay for their own complexity. It is the purest version of the personal-site project — a constraint nobody asked for, an acceptance criterion of 'recognizable enough', and a written apology to Nepal.

New York & the Table

Eater NY

The Public Theater's New Restaurant Pays Homage to Old New York

The Astor opens today in the Public's old Library space, wood-paneled and pitched at the taverns and supper clubs the building outlived, with Andrew Carmellini and the Joe's Pub team behind it and a Bar Torino café alongside. The trend underneath it is the interesting part: theaters and museums have stopped treating the restaurant as an amenity and started treating it as the reason people arrive early.

Hell Gate

'This Is the Most Political Film I've Ever Made': 15 Minutes With John Wilson

Wilson's feature debut, The History of Concrete, opens Friday and covers Elizabeth Street Garden, the Bed-Stuy Aquarium, the colonization of Mars and the problem of gum on the sidewalk, which is roughly the How To method at feature length. He shot ninety-nine percent of it himself after failing to get it financed, and it walks straight into the housing-versus-preservation fight his earlier work kept circling.

Loose Change

Aeon

The curio cabinet of medieval mental health

The demonic-possession stories are real and they are also the wrong sample: Katherine Harvey went looking outside the saints' lives and found physicians prescribing exercise, gardens, animals, company and music alongside the saffron and ostrich liver. What the Middle Ages lacked in pharmacology it made up for in an assumption we have spent the last century relearning — that a mind is treated by changing the life around it, not only the body inside it.

The Back Page

The Guardian

VARs who wrongly confirmed Haaland's goal at Manchester United are stood down

Matt Donohue got the semi-automated offside call right and then missed Enzo Fernández influencing play from an offside position, so Webb has left him off this weekend's appointments — the first standing-down of the season. The machine resolved the part it was built for and the judgment call underneath it still decided the Manchester derby, which is the shape of every automation argument at work.

Around town

Wed Sep 16 · 7:00am

Avalanche Summit

Chelsea Industrial, 535 W 28th St, Chelsea

Bitwise's Matthew Hougan and Franklin Templeton's Sandy Kaul on the same bill as Meltem Demirors and Haseeb Qureshi, which is the rare crypto lineup where the allocators and the traders have to argue in front of each other. A hundred dollars, and the institutional-adoption talk track is his day job's weather report.

professional

Wed Sep 16 · 6:00pm

LeadDev After-Hours

Coexist GameHouse, 325 W 38th St, Garment District

The unofficial second night of LDX3, in an arcade rather than a ballroom, hosted by Rootly, Docker, Runpod and friends. Free, and the conference hallway track is reliably worth more than the stage — this is just the hallway with better lighting.

casual

Thu Sep 17 · 7:30pm

Antimetal Dinner

Roscioli, 43 MacDougal St, SoHo

A curated table of technical leaders from fast-growing startups, at the Manhattan outpost of the Roman deli, which is a better setting for an honest conversation than any rooftop mixer. Free and small, so the vetting is the product.

professional

Fri Sep 18 · 5:00pm

Space Apps NYC: Exploring This Year's NASA Challenges

NYU Tandon, 370 Jay St, Downtown Brooklyn

The kickoff for NASA's global hackathon, where the challenges are real open datasets and the deliverable is a weekend build rather than a pitch deck. Free, and a reminder of what building for fun felt like before it had a roadmap.

casual

Wed Sep 16 · 7:30pm

Kwaidan (Masaki Kobayashi, 1964)

Film Forum, 209 W Houston St, West Village

Four Japanese ghost stories told across three hours of hand-painted skies and Tōru Takemitsu's score, made by a director who built his weather indoors rather than wait for it. It runs through Thursday only, and it is the kind of film that collapses to nothing on a laptop.

casual

Quote of the day

“'Delete the temp files' and 'delete the wrong files' are one fat-fingered glob apart.”

— Daniel Botha, Fly.io

On this day

September 15, 2008 — Lehman Brothers files for bankruptcy in the Southern District of New York, $613 billion in debt, the largest filing in American history and still the one every risk model quietly measures itself against.

The puzzle corner

The Mini · Wordle · Connections

Four months apart, Anthropic and Fly.io arrive at the same answer: stop trying to watch the agent, and fence the room it works in.