Telemetry

How to leave your laptop at home and do more than you ever did.

Jeremy Smith · An essay
OpenDraft · not signed off

Phone or laptop

There is a man called Boris Cherny, and if you write software for a living you have probably heard his name said in the tone people save for someone who must know something they don’t.

Boris made Claude Code, more or less by accident. He had joined a small team inside Anthropic whose job was to find things the models could already do that nobody had built yet, and in late 2024 the state of the art in AI coding was pressing tab to finish a line. He thought the model could write the whole thing. The first version, as the story is usually told, came together in about two weeks. By his own account it then didn’t really work for six months. He used it for maybe a tenth of his own code. It launched, and it was not a hit. Then the models caught up with it, and it was.

Earlier this year he sat on a stage in front of a room full of people who build things. I like to imagine two developers near the back, the kind who arrive early for the seats nearest the power points.

Before the interviewer could get going, Boris asked the room a question of his own. Who uses it mostly in the terminal?

Most of the hands went up, both of theirs included.

“Obviously,” whispered one. “He’ll be terminal.”

“Forty panes,” said the other. “Minimum.”

The desktop app?

“Nah.”

Inside their code editor? Not many, which seemed to surprise him.

“He doesn’t need an editor,” whispered the first. “He is the editor.”

Other?

“Hundreds of terminals,” said the second, leaning in. “Thousands. I’ve heard things.”

Then Boris answered his own question. “I’m like iOS mostly these days.”

The room laughed. The two at the back laughed loudest. It was a good joke.

Later he was asked about his setup, and he held up his phone. The Claude app, a code tab, five to ten sessions. Inside those, a few hundred agents. Overnight, a few thousand. “You probably can’t see it,” he said.

The two at the back had the number right. They had everything else wrong. One of them looked down at the laptop on his knees, which was open, warm, and running four terminals, the way you might look at a horse after someone has driven past in a car.

I see that laptop every day. I make coffee for a living, and the café I work in is full of people building things on laptops. They come in at nine, take the table by the power point, order a flat white, and open a chat window with a model in it. Then they sit there. They type, the model types back, they read what it typed, they type again. Around eleven they order a second coffee without looking up. The first one is still on the table, mostly full, and quite cold.

I don’t say anything. It’s how I started. But it’s a conversation, and a conversation needs both people in the room.

I can’t be in the room. I’m on the other side of the counter with a milk jug in one hand and a portafilter in the other, and a queue that doesn’t care what my agents are doing. If anything I build is going to get built, it has to get built while I’m doing something else, and the only thing I can look at between coffees is my phone.

So I learned to work a little more like Boris and a little less like the people by the power point. Not because I’m hardcore. Because I didn’t have a choice.

So this, in order, is how you fail repeatedly to run a boatload of agents.

The first time, it went perfectly.

Open

Source: a Sequoia-hosted fireside, 2026 (YouTube SlGRN8jh2RI). The poll, the iOS line, the six months that didn’t work, the tenth of his code, “not a hit”, the phone setup and the numbers are all his words there. The two-week build is from secondary sources, hence “as the story is usually told”. The whisperers are invented and marked “I like to imagine”.

About 620 words with the café transition, against a 2,000 target. The transition is first person; the Sydney night that follows is still “a man in Sydney”. Third person throughout, or “I” from here on?

IDraft · not signed off

Built to fail

On the night of the thirtieth of July, a man in Sydney went to bed and left a computer in charge of a website.

It was a small website for a small shop, and the man had been building it one feature at a time for weeks. He now believed, with the confidence of a man who has just written a very long prompt, that it could build the rest itself.

The arrangement was elegant. One agent planned. A second built what the first had planned. A third, which had never met the second and was forbidden to read its reasoning, checked the work cold, like a building inspector who is not allowed to speak to the builders and communicates only by leaving notes under the door.

By morning the machine had attempted seven features. Four passed inspection first time: a gift-card balance checker, a delivery-time estimator, a loyalty stamp card, and a page showing what was in season. The other three it had politely declined, for reasons it had written down and which turned out to be correct. Nothing had needed redoing. By every measure the night had set itself, it was flawless, and the man put all of it live before lunch.

Four weeks later, not one of the four was known to work.

Two were waiting on something nobody had noticed: the shop had already used every discount code its account allowed, and a stamp card that hands out discounts can’t do much about that. One had simply never been opened. The fourth had been opened by nobody either, but a later look at the code showed it was cheerfully presenting yesterday’s menu as today’s. The fix was a single character.

Nothing had gone wrong, exactly. The loop had a rule, written in advance and quite sensible: the machine may say the work passes, but only the man may say it works. It followed that rule perfectly, and then it waited. It had been built with no way to tell him what was waiting, or which thing needed him, or why. Four perfect results sat in a queue marked “in progress”, which is the status things have while they quietly go off.

The loop had been built to fail. Not loudly. Politely, and in the dark.

This is an essay about turning the lights on.

Open

“Went to bed” is an assumption; the record says only “unattended”. Possible true line: “The real one was a game add-on, and it ended when my character died.”

Honesty line still to place, e.g. “The details are changed. The shape isn’t.”

IIDraft · not signed off

A bench

The second time, before he let the machine build anything at all, he had it build him a bench.

A bench, in the sense he meant it, is not furniture. It’s the thing a craftsman puts the work on so he can look at it properly. This one was a page. It listed every feature the machine had built since he last looked. Next to each was a button that took him straight to it, already set up in the state where it was meant to be tested: a gift card with money on it, an order placed four minutes before closing, a loyalty card one stamp short of a free coffee. Under each was a box, and the box asked the only question the first loop had never asked him: well?

He could click through all of it on the bus. He could say “the stamp animation is too slow” or “this says Tuesday and it is Wednesday”, and that sentence didn’t go into a void. It went to the builders as their next instruction, and by the time he got home it was usually dealt with.

The first loop had treated his judgement as the end of the line, the place where finished work went to wait. The bench made it the start of the next lap.

It also taught him something about instruments, which is that they lie. Not out of malice. A check that can only see part of a thing will report on that part with total confidence. One of his early benches announced, quite firmly, that a button didn’t work. He pressed it. It worked. After that he had a rule: a readout that says “I don’t know” is useful, and one that says “no” when the answer is yes is worse than having no readout at all.

[One line on what the bench changed: features verified per week, or how fast critiques came back fixed. Not invented; no number is in the store yet.]

Open

“On the bus” fits the subtitle. If the real bench was a desk thing, move the phone beat to the picks section.

IIINot written

The chat app was the wrong tool

To write

Hooks first: code the harness runs on an agent event, whether or not the model remembers to. A chat app has nowhere to put one. Coding agents over chat apps, directed from the phone: tool-call ceilings, sessions dying when the phone sleeps, nothing running while nobody watches. It’s about the harness, not the model.

IVNot written · figure ready

Which one is which

To write

Tier 1, no build needed: ask your agent for a hook that retitles the session as the work changes. Titles are for routing (which chat do I tap), and a few characters are enough when you know what they mean. The figure needs a caption that decodes them.

BeforeTitles set by the first prompt
AfterA hook rewrites them
The same agents doing the same work. Before: every title is whatever the first prompt was about. After: a hook rewrites each title when its agent changes task, and sub-agents appear under the session that spawned them. Names and tasks are invented.
VNot written · figure to build

Where is everything up to

To write

Tier 2: a feed. Heartbeats, status counts (working / idle / quiet / ended), full titles, permission level, what needs you. Absence isn’t visible, so “nothing happened” looks like a quiet day. The heartbeat hook that failed silently: telemetry needs its own telemetry.

Figure: SPINE feed, anonymised. Status counts and a few rows.

VINot written · figure to build

It needs my taste and I’m not there

To write

Picks: the recommendation already chosen so agreeing is one tap, evidence attached, honesty labels (“Model only”), a note, Skip. Taste arrives as a batch of rulings, not a conversation.

Figure: picks card, anonymised, from the batch pattern.

VIINot written · figure to build

Send is a hook too

To write

Hooks come out of the agents so you can see them; a hook goes back in so they start again. And the catch for when it doesn’t: approved, not done.

Figure: the picks mock ends on Send, then cuts to the session list waking with new titles.

VIIINot written · figure to build

Keep them spinning

To write

“As soon as one is blue I’m voice prompting.” Signal, label and a cheap reply. Then the close: snowball, then avalanche.

Figure: the title-hook mock extended: a row goes blue, a voice prompt is sent, it spins again, the hook rewrites the title.

Still needed: one real number for the subtitle’s “more than you ever did”.

Set in Instrument Serif and Newsreader · Working draft, not published