The First Hour: transcript

[Incident-2277] Ep 1 - Idle Is Not Done

Cold open

Kit (sleepy, then alert): Two oh seven. Okay. Okay, I'm up. Page says: checkout slow, the slowest checkouts are ten times slower than normal.

Kit (quiet, reading): And there's Nell's sticky note on my monitor. You're primary. I'm on leave. Don't call me. When it goes quiet, read why.

Kit (quiet): Not calling you, Nell. First solo week. I've got this. Probably.

Act 1

June: Kit? June Park, customer operations. I'm on the overnight shift, and my support queue just lit up. People can't pay for their groceries. What's going on, in words I can use?

Kit: Hi June. When somebody presses pay, it's normally a second or two. Right now the slow ones are taking most of a minute, so people give up. I don't know why yet.

June: And the robot? I was told we have a robot now.

Kit: Kind of. An AI agent. A program that reads our logs and asks questions on its own, so a person doesn't have to start from zero at two a.m. Nell built it before her leave. It's been on the case since the page.

June: What's it doing?

Kit (reading): I can watch its trail. Three kinds of steps. A send is our app handing it the problem. A watch is it telling us what it's doing, we just listen. An answer is it asking our app for something only we have, and waiting for our reply.

Kit (reading): So. Send: checkout is ten times slow, what happened. Then watches. It's reading the checkout log. Lots of customer email addresses scrolling past in there, by the way.

June: Is it allowed to read those?

Kit: Reading is allowed. It can type commands to look at files without asking anyone. Changing anything is a different story. Okay, now three answers. It's asking our app for the list of recent changes, and for two sets of graphs.

Kit: And then. It went quiet. Our app wrote one word. Done.

June: Done? My queue is still filling up. The six a.m. delivery slots lock at two thirty. If people can't pay before then, a lot of families don't get groceries tomorrow morning.

Kit: How many stuck?

June: About eleven hundred carts.

Act 2

Kit: I'm escalating. Marcus is on the escalation rota this week.

Marcus (level, brisk): Marcus. I'm here. Kit, short version.

Kit: Checkout ten times slow since two oh seven. The agent read the logs, asked us three questions, went quiet. Our app marked it done. Eleven hundred carts stuck. The delivery window locks at two thirty.

Marcus: Four Thanksgivings ago a script restarted our payment system in the middle of a sale, because nobody was watching it. I agreed to this agent because it can't restart anything without a person. Now it's doing nothing, quietly. Wonderful.

Kit: Nell built it. She'd know in two seconds. I could just...

Marcus (firm, kind, at a normal pace): Nell is on leave. We don't call people on leave. What does her note say?

Kit: When it goes quiet, read why.

Marcus: Then read why.

Tag

Kit (reading): Okay. The quiet moment itself has a reason attached. Status: idle. Stop reason: requires action. And three ticket numbers.

June: In plain words?

Kit: Idle just means it stopped talking. The reason says why. If the reason said end turn, it would mean finished. Requires action means it's waiting on us. And those three ticket numbers are its three questions. Nobody answered them.

June: Oh, I know this one. In my world that's a ticket set to waiting on customer. It isn't solved. It sits there until the customer writes back.

Kit: Exactly. And the customer here is us. Our app saw a quiet ticket and marked it solved. Every quiet. Finished or waiting.

Marcus: Then answer the agent. You have twenty-one minutes.

[Incident-2277] Ep 2 - Who Answers

Act 1

Kit (fast): Two twenty-one. Nell's runbook has a page for exactly this. If it's waiting, run resume. Resume finds every question still open, gets our answers, and sends each one back with its ticket number, so the agent knows which question each answer belongs to.

Marcus: What does it touch on our side?

Kit: Three lookups. The list of recent changes and two sets of graphs. Read only. Nothing changes.

Marcus: Your call. You're primary.

Kit (steady): Running it. Three answers sent. It's moving again.

Kit (reading): It says: a change to the payment code went live at one fifty-two. Payments started waiting at one fifty-three. Now it wants one more thing from us. The diff for that change.

June: The what?

Kit: The exact lines somebody edited. Like track changes in a document.

Kit (reading): Answered. And there's one line that matters. Pool size. Fifty, changed to five.

June: Which means?

Kit: Pool size is how many lines we keep open to the card processor at once. Think of checkout lanes at a store. Somebody closed forty-five of the fifty lanes, at one fifty-two in the morning, and the line out the door is every cart in your queue.

Kit (quiet): And the person who approved that change on Friday at five p.m. was me.

Marcus (firm, warm, at a normal pace): Kit, stop for one second. Nobody is in trouble. We run a blameless postmortem after every incident, and this one will be blameless too. We'll ask what let that number ship, not who. A tired reviewer on a Friday is a fact, not a cause. That's a problem for later, and later it's a problem for the whole team, not you.

Kit: Okay. Okay. Thank you.

Marcus: Tonight we fix checkout. Keep going.

Act 2

Kit (reading): It went quiet again. Requires action. One question this time. It wants to restart the payment service.

June: Restart, like turning it off and on?

Kit: Exactly like that. And it can't do that on its own. That request comes to our app, and our app asks the person on call before it does anything. Which is me. My phone just buzzed. The agent wants to restart payments. Yes or no.

Marcus: There it is.

Kit (fast): Two twenty-four. A restart takes forty seconds. Undoing the change, a rollback, takes six minutes. The window locks at two thirty. Marcus, what would you do?

Marcus: I'd want the person with the most information to decide. That's you. You have the diff open.

Kit (quiet, to themself): Nell would know. Don't call Nell. Read why.

Act 3

Kit (slow): Okay. A restart turns the service off and on with the same settings. The same five lanes. It would look better for a minute, then the line would build right back up, and we'd have restarted payments in the middle of the delivery window. Which Marcus has a story about.

Marcus: I do.

Kit (decisive): So, no. I'm answering with a no and the reason: the change cut the lanes to five, a restart keeps five, the on-call engineer is undoing the change instead. That goes back to the agent as its answer. And I'm starting the rollback myself. Now. Two twenty-four.

June: What do I tell my team?

Kit: Tell them six minutes. Then tell them to try the stuck carts again.

Kit (reading): The agent read my no. It's writing its summary. The on-call engineer declined the restart, because the settings would come back with five lanes. Recommended fix: undo the change from one fifty-two. And the stop reason this time is end turn.

June: End turn. Plain words?

Kit: Finished. It said what it found, and it stopped because it was done. Not because it was waiting.

Tag

Kit (quiet): Two thirty. On the second. Rollback complete. Errors falling.

June: Carts are going through. My team can phone the ones that missed the window. About sixty.

Marcus: Nice job, Kit, and thanks for respecting Nell's leave. Go back to sleep. You can write up the report tomorrow.

[Incident-2277] Ep 3 - The Stop Table

Act 1

Marcus: Morning, everyone. I hope you had a good weekend. First, welcome back from leave, Nell. Let us know what you need as you ramp back up. Alright, Kit, let's get into the postmortem. Blameless, as always. Facts first.

Kit (calm and unhurried, rested, a measured daytime pace): Two oh seven, page. The agent read the logs and asked our app three questions. The app took the quiet to mean done, so nobody answered. Two twenty-one, I answered by hand. The agent found the change that cut the payment lanes from fifty to five, and asked to restart payments. I said no and undid the change. Two thirty, recovered. Sixty carts missed the window, and June's team phoned every one.

Nell: And the app that took quiet to mean done is mine. I wrote that line. It treated every quiet the same, finished or waiting. That's my fix to make.

Marcus: Before we get to action items. Rosa, you asked to join this one.

Rosa: I did. And I'm sorry, it's the boring question, but it's my job to ask it before we decide anything. The agent read our checkout log. Where does what it read live now?

Nell: With the session, on Anthropic's side. That's what a managed agent is: Anthropic runs the agent for us, and they keep the whole transcript there: everything it was told, everything it read, every answer.

Rosa: Then it isn't eligible for zero data retention, the promise that data is deleted right after it's used. Their overview page says so plainly, because keeping the history is how the service works. And that log is full of customer email addresses.

Rosa: And there's a second part. Our privacy policy tells customers how we handle their data and who we share it with. Sending checkout logs to an outside service may or may not be covered by what we promised. I don't want to guess. I want to check with legal before the agent reads anything with a customer's details in it.

Marcus: So the agent's job is on the table.

Act 2

Marcus: Option one. We run the whole thing on our own servers, with their agent software. The transcript stays with us.

Rosa: Stored with us, yes. But every time it asks the model a question, the conversation still goes to the model.

Nell: To be clear, option one isn't a managed agent anymore. It's their agent software running on our machines. And then at two a.m. we're the ones keeping that machinery alive. The exact job we built this to stop doing.

Kit (calm and unhurried, rested, a measured daytime pace): Option two. Keep their service, but put the scratch computer it works in, the sandbox, on our side. The files stay with us.

Nell: That one is still a managed agent. Anthropic still runs the agent. Only the sandbox moves to us.

Rosa: But whatever it reads in those files still travels back to their side so it can think about it. Their docs say that too. Where the files sit isn't my worry. What it reads is.

Marcus: Then we choose by what we stop owning. We stop owning the machinery at two a.m. We keep owning the data. Rosa, we'll need to scrub the customer emails from the log read by the agent.

Rosa: Mine. Done by Friday. And I'll take the privacy policy question to legal this week. Until both are settled, it reads graphs, not logs.

Nell: So we keep the managed agent. And the rules, in one breath. Reading is fine with nobody watching. It doesn't type a command or restart anything until the on-call engineer says yes. And it stops at two dollars.

Kit (calm and unhurried, rested, a measured daytime pace): Two dollars?

Nell: A spending cap. When it hits two dollars it stops, and nothing we send can wake it back up. So the app can tell us if it's exceeded its budget, instead of going quiet.

Kit (calm and unhurried, rested, a measured daytime pace): Then here's what our app does every time the agent goes quiet. We'll call it the stop table. The agent always says why it stopped, and each reason gets one response. If it says it's finished, the app shows its answer to whoever is on call. If it's waiting on us, the app answers every question it asked, matched up by ticket number, so none gets skipped. And if it ran out of budget, the app reports that and stops, because nothing we send will wake it back up.

Marcus: And mine. An alert if it's ever waiting on us for more than two minutes. I've had a rule about restarts for four years. I need one about silence.

Act 3

Kit (whispering): Tuesday, one week later. I'm primary on call again, and it's three twelve in the morning. Everyone in my apartment is asleep, so I'm whispering. Checkout errors are climbing again.

Kit (whispering): Phone says the agent wants to type a command to read the cart log. That's the new rule: no commands without the on-call's yes. Yes.

Kit (whispering): Three nineteen. It wants to restart the cart cache. Memory full since three oh one, and nothing has shipped since Friday, so this time a restart is the right fix. Yes.

Kit (whispering): Three forty. Message from the app. The agent stopped at its two dollar cap while it was double checking. It can't be woken, and the app said so instead of going quiet. Errors cleared at three twenty-one. Report attached. Going back to sleep.

Act 4

Nell (bright): Nine a.m. Kit, I read your report with my coffee. Clean. Nobody woke up who didn't have to.

Marcus: I read it too. One line in particular. Three nineteen: restart of the cart cache, approved by the on-call engineer. Four years ago a restart happened because nobody was asked. This one happened because somebody was. Same button. I'm glad to see we're maintaining that rule.

Kit (calm and unhurried, rested, a measured daytime pace): Your note's still on my monitor, Nell.

Nell: Keep it. It works on everybody.

Tag

Kit (warm, clear, unhurried, a relaxed Friday pace): Friday. I'm updating the playbook, and recording the handover for Priya, who's on call next week. Priya, five things.

Kit (warm, clear, unhurried, a relaxed Friday pace): Number one. The agent lives on Anthropic's side, so nobody here keeps that machinery running at two a.m. We still own the data: the logs it reads are scrubbed, and legal has signed off on what we share.

Kit (warm, clear, unhurried, a relaxed Friday pace): Number two. Our app sends the agent the problem, watches what it does, and answers when it asks for something only we have.

Kit (warm, clear, unhurried, a relaxed Friday pace): Number three. When it goes quiet, read why. Finished means show the answer. Waiting means answer every question by its ticket number. Out of budget means report it and stop, because nothing wakes it back up.

Kit (warm, clear, unhurried, a relaxed Friday pace): Number four. Reading is fine with nobody watching. Commands and restarts wait for a yes from the on-call engineer, and that's you. If the evidence says no, say no and say why.

Kit (warm, clear, unhurried, a relaxed Friday pace): Number five. Nobody on leave gets called. Escalate to the rota. And if something you approved breaks, it's a blameless review in the morning, not a confession at two a.m.

Kit (warm, clear, unhurried, a relaxed Friday pace): Good luck. You've got this.