Anthropic has published how it used a Claude chaos loop to get Claude Docs ready for launch: an AI agent whose whole job, for two weeks, was to break the service on purpose and then fix what broke. The result was 60 server errors in about 2 billion requests and not a single rollback. Netflix made Chaos Monkey famous more than a decade ago. This version doesn’t just pull the plug; it reads the logs, writes the fix and asks before touching anything it can’t undo. That last part is the bit we think every team building with AI should copy.
How the Claude chaos loop worked
Claude Docs is Anthropic’s shared document editor, where people and Claude write and comment on the same doc live. According to Anthropic’s engineering write-up, three weeks before its public launch on 16 September it was still a single-node proof of concept used by a few thousand people inside the company. Six days after launch, people had created more than 4.6 million docs.
In between, the team ran a loop on a staging copy of production:
- A test harness picked a scenario and pushed load until something gave: 10 million new docs in a day, 22,000 connections reconnecting at once, a server killed halfway through loading a doc, a stalled database.
- Claude read the telemetry, proposed a short menu of fixes with a recommendation, and for changes that were easy to reverse, wrote the fix, deployed it to staging and ran the test again.
- A separate Claude session reviewed every change before it went in.
- Anything hard to undo, like the database schema or how docs are stored, needed a human engineer’s approval.
Claude also had two promises it was never allowed to break: docs open almost instantly, and once the server acknowledges an edit, the edit is kept. The best catch was a single large doc using about 3 GB of memory, because every keystroke copied it into several data structures. Fixing that took the estimated launch capacity from more than 500 pods to about 50. Seven of the ten scenarios rehearsed on staging then happened for real in launch week. The inference bill was about $125,000, against what the author guesses would have been six months of work for a team of engineers.
Our version: let the strict client find it first
We don’t serve billions of requests, but we have a smaller cousin of this loop running right now. We are building native iOS and Android apps for Calendrz, and most of the code is written by AI agents working from a written plan. Our founder, Liviu Tudor, split that plan into numbered steps, each one a separate agent session, with a gate between stages that says what must be true before the next one starts.
The step that behaved most like a chaos loop was the API contract. The backend now publishes a machine-readable description of every endpoint the apps call, and the apps’ network code is generated from it. Generated clients are wonderfully unforgiving. The web app had quietly tolerated a few loose ends; the generated clients refused them straight away:
- A field the web app’s “Poll started” card relied on was described in the contract, but nothing ever filled it in. Fixing it for the apps made the card show again on the web too.
- One refusal answered with the status code that means “your session has ended” rather than “you can’t do that”. Both apps read that code as a sign-out, so a harmless mistake would have logged people out.
- A field the contract promised would always be there could, in rare cases, be missing. A strict client can’t read a reply like that at all, so the backend now leaves the whole section out instead of sending half of it.
None of these would have crashed anything on the web. Each would have been a confusing bug report from someone’s phone. Each fix came with a test that checks the contract, and each went through a separate review round before merging, just like Anthropic’s second Claude session.
The gates are where people stay
The other parallel is the line between what an agent may do alone and what waits for a person. In Anthropic’s loop, reversible fixes went straight to staging and irreversible ones waited for an engineer. In our plan, agents write and test the code, while the steps that are hard to take back, like creating the app store records or the keys that sign each release, are marked for a human and never handed to an agent. That is the same instinct behind availability mirroring itself: auto-blocking your time in every connected account is safe to automate, because a block is easy to move or remove. Booking a meeting on your behalf is not, which is why our Smart Scheduler proposes a time and waits for you to say yes.
What to take from it
We believe software is heading towards a “lights out” setup, where people write the specs and the promises, and agents write, test and ship the code. The chaos loop is a good picture of what that looks like when it works:
- Write the promises down first. “An acknowledged edit is never lost” is something an agent can defend for weeks. “Make it solid” is not.
- Let a machine attack it before your users do. Load scripts, killed servers and strict generated clients all find the bugs that polite testing misses.
- Gate what you can’t undo. Speed on the reversible, a human on the irreversible.
And if you’re choosing tools that touch your calendars, ask their makers what happens when things go wrong: a meeting moved twice, a calendar that goes offline, an account that needs signing in again. Our calendars already bring enough chaos; the software shouldn’t add more 🙂
Want your availability to stay right across every calendar you juggle? Connect your calendars to Calendrz and let auto-blocking keep you from being double-booked.







