Logging in again was wrong: a morning with a provider outage
Two of my coding sessions got the same error: log in again. The status page said something else. Seven rules from a provider outage that lasted two hours.

TL;DR
On 13 November 2026 two of my coding sessions got "Please run /login" between 11:16 and 11:19. Anthropic's status page showed login itself broken since 10:58, all-clear at 12:58. Logging in again would have fixed nothing. Seven rules: status page first, accounts untouched, one voice, counting over guessing, an alarm over watching, diff after a stop, separate loads.
Friday, 11:16, two sessions stumble
The error said: log in again. The right move was not to.
Friday morning in Vienna. My AI agents are working, each in its own coding session, spread across several repos. Above them runs a lead session. It writes no code. It keeps track of the whole picture, tells the others what the rules are right now, and collects what I need to decide.
Between 11:16 and 11:19 two sessions get the same message, four times each:
Please run /login · API Error: 403 Attention Required! | Cloudflare
Both had only been started that morning. The message is polite and clear. It tells you what to do: log in again. An agent that takes instructions seriously does exactly that.
Around 11:20 I notice something is off. I tell the lead session there are server problems. I ask it to look into it and set an alarm.
The answer was already on the status page
At 11:22 the lead session reads seven status pages. Anthropic has had an incident open since 10:58, severity "major", status "investigating", titled "Elevated errors on claude.ai, Claude Code, and platform.claude.com". The text below it: a problem with logging in to Claude.ai and Claude Code.
That fit the message. The incident had been there since 10:58, before the first errors I counted. The login itself was what had broken. Logging in again would have fixed nothing. At best it would have failed at the same spot. At worst it would have touched accounts in the middle of an outage.
The rest of the sweep was calmer. OpenAI reported a minor incident that only concerned a connector. Cloudflare reported a minor one as well. GitHub and Vercel had no incident, the Azure feed had no entry, and my own GitLab answered with 200.
That is the first thing I take from the morning: status page first, own diagnosis second. An error message describes what the client sees. What is going on at the provider is written somewhere else.
Nobody touches the accounts
At 11:25 both sessions get the same short instruction. No /login. No account switch. Retry the step after two minutes. Start subagents one after the other.
That is the second rule. While login is broken, nobody touches the accounts. No login, no switch to another account, no new sessions. And no handover of the lead session to a successor, because the successor depends on login too. If something goes wrong on a morning like this, I want the outage to be the cause. A change I start in a hurry only adds a second one.
One voice makes the call
Both got the same wording from the same place. If every session guesses on its own, you get as many answers as sessions. Sooner or later one of them is "I'll log in again". That is exactly what the error suggests.
The lead session works like this on ordinary days too. How the fleet around it is built is in Building an agent fleet: Sven, BuilderBob and their loop. That morning the role paid off. The right answer was one no error message points to: leave the accounts alone and retry the step later.
Eight counted, two not
At 11:23 the lead session counts. It reads the session transcripts and counts the errors per session, with a time window. Eight sessions were covered. Two of them had the errors from earlier, six kept going without a single API error. Two more sessions were not covered.
I do not count those two as healthy. "Not covered" means I don't know. Counting them as "not affected" turns a gap in the measurement into good news.
An alarm instead of staring
I could have reloaded the status page myself every few minutes. All it took from me was one sentence. The lead session set the alarm itself, and it ran from 11:25. Every five minutes it checks what the status pages say. It only reads, it changes nothing. It speaks up when something changes, and it ends itself once the all-clear comes.
That way I did not have to keep an eye on the page myself. An alarm that shouts "still broken" every five minutes would be background noise within the hour. I measured what a warning that almost always fires is worth, on a RAM warning: A signal may only warn if it is rare.
The helper that stopped halfway
One of the two affected sessions kept working, with subagents one at a time. One of them stopped with exactly this 403, in the middle of its work.
The session then read its diff. It showed three files the subagent had deleted. Whether that was half a step or intended, a helper that stopped can no longer tell you. The three files have been restored.
A helper that stops can leave a half-finished state behind. The stop only tells you that something ended. Only the diff tells you what happened up to that point. That holds on days without an outage too. Why I read what agents produce as a matter of course is in Critical thinking when AI agents do the work.
What did not depend on the broken service kept running
In the same hour a fair amount ran without any trouble. From 11:07 to 11:11 my CI server restarted with a new kernel. From 11:15 to 11:17 a merge of this website went live, via GitHub and Vercel.
That came from separation. Whatever does not depend on the affected service carries on as usual: servers, deploys. Whatever does depend on it runs leaner, meaning subagents one after the other.
What I don't know
There are a few things I don't know. They are in here too.
First, the gaps in my measurement. Two sessions were not covered. Whether they had errors, I don't know. In the covered sessions, errors only occurred between 11:16 and 11:19. From 12:01 to 12:59 the five-minute check found not a single one in the covered sessions. I did not measure 10:58 to 11:16.
Second, my products. Some of them have AI features that call models through API keys. According to the status page, claude.ai, the Console, Claude Code and Cowork were affected. Whether those calls were affected too, I have not measured. Before I write anything about it, I read the error rate for each product. Until then there is no statement here, not even a reassuring one.
The evening before, there had already been a different incident. At 18:52 Anthropic reported elevated errors in the Microsoft 365 connector, at 19:29 OpenAI posted the same title. At 23:23 Anthropic named an incident at Microsoft 365 as the cause. My fleet does not use that connector, and I have not measured any effect there either. It is a separate incident, and at 12:59 it was still open.
According to the status page, the login outage ran from 10:58 to the all-clear at 12:58, about two hours. Anthropic says a fix was rolled out from 12:05, and the time after that was monitoring. The text of the all-clear names claude.ai and Claude Cowork, but the whole incident, with Claude Code in its title, is marked resolved. Up to the all-clear Anthropic named no cause, and I am not guessing one.
The message was not wrong
What stays with me most from this morning is the error message itself. It was not wrong. Had my login expired, "log in again" would have been exactly right. That morning, though, the login itself was broken, and an agent that only reads the message cannot tell the difference. The status page can.
That is why it now comes before the diagnosis. The other rules joined that same morning, each with its reason. How I set up processes like this for several agents is under Foundation. If you run several coding sessions in parallel yourself and are working out how to handle outages, write to me.
lead-session-orchestrator
Many sessions, one lead, and whatever only you can decide waits in one inbox.
- The language model
- runs the lead session: it reads the situation, orders the merges and bundles your decisions.
- The code
- sets the limits: only ever one lead, one rule file per repository, and whatever it cannot measure, it leaves alone.
session-orchestrator.com/lead (opens another website)Product page for lead-session-orchestratorWhat I build for you