Since 14 August, new Claude Code sessions on the Pro, Max and Team plans start in auto mode, exactly as Anthropic said they would. No surprises there. The more interesting things happened beside it: Anthropic began watermarking the text Claude generates, and on Sunday evening the whole of Claude went quiet for half an hour.
Claude starts watermarking the text it writes
The text you get back from Claude will carry a watermark – an invisible pattern woven into the words themselves. It is not a note appended at the end, and not metadata you can strip in an editor: the pattern sits inside the text and travels with it when you copy and paste it somewhere else. In time it will cover Claude, Claude Code, Cowork, the API and Claude Tag, which is every route to the model there is.
There is a catch that is easy to miss, though. Models released on or after 2 August 2026 carry watermarking from launch, and Anthropic is still building it into the earlier ones – except every model you can work with today came out before that date. Sonnet 5 and Opus 5 arrived in July, Fable 5 in June, Haiku 4.5 last year. Nothing you wrote with Claude last week has a watermark in it.
The mechanism is surprisingly simple. While writing a response, the model repeatedly faces a choice between words that fit equally well, and normally settles it at random. Watermarking replaces that randomness: the choice follows from a key and the handful of words before it, which leaves a pattern across the whole response. Anthropic says a watermarked answer and an unwatermarked one are indistinguishable to a reader, and that response quality doesn’t change.
What the watermark does not do is a separate matter, and one where confusion has already piled up. There is nothing in it that identifies who was writing – no account, no session. It doesn’t tell Claude apart from another AI model either; it says only that Claude was probably involved somewhere along the way. On a short passage it won’t work at all, because there are too few choices between words to leave a pattern. It also thins out in factual writing, where the model has less room to move, and doesn’t apply where the output has to be exactly one thing and not another – which means code. Light editing probably won’t remove it; rewriting the text word by word will.
Nothing in your daily work changes today, which is the honest summary. There will be no way to switch the watermark off, and there is no detection tool yet – Anthropic promises a detection API “soon”. So what has been announced this week, with the full apparatus of an EU obligation behind it, is for now a feature that doesn’t exist. Our reading is that this follows from the regulation rather than from anyone asking for it. And if you publish text written with Claude, the practical conclusion is older than any of this technology: say plainly that you use AI, instead of counting on nobody checking. That is cheaper than any strategy built on pretending otherwise.
What’s new in Claude Code
Auto mode is the default now, and Windows had a rough day
From 14 August, Claude Code starts in auto mode if you are on Pro, Max or Team and haven’t set anything else. In that mode the AI assistant doesn’t stop before each command – a separate model judges whether an action is safe enough not to ask about. What that means in practice, and which mode suits which stage of your work, we take apart in our tip on permission modes.
Version 2.1.232, from 13 August, introduced a bug along the way that hit Windows users only: auto mode kept stopping for manual approval on ordinary commands, the kind that change directory and write output to a file. Version 2.1.233 fixed it a day later, and Anthropic rolled back the changes behind it. So if you work on Windows and the new auto mode looked broken to you on Thursday, it wasn’t your imagination.
The task list is gone from newer models
The task-tracking tools are gone from every newer model – Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and whatever follows them – as of version 2.1.233 of Claude Code, released on 14 August. This is the mechanism Claude Code used to break longer work into steps and tick them off as it went – the one that printed the list of what it was about to do on your screen. You can bring it back by setting the CLAUDE_CODE_ENABLE_TODO_TOOLS=1 environment variable.
Anthropic doesn’t explain the removal in the changelog, and we won’t guess at it. What we will say is what gets lost. The task list was the only place where a long stretch of work showed its progress: how many steps are behind you, how many are left, and which one the AI assistant is standing on right now. Without it the screen shows a stream of whatever is happening this second, with nothing to tell you which part of the job you are looking at. It did one more thing on the side – drawn up at the start, it revealed whether the task had been understood the way you meant it, before anything ran.
If you work with Sonnet or Opus, asking for a plan won’t give that view back: a plan tells you what Claude Code intends, but it doesn’t update as the work goes. Only the environment variable above restores it.
Which model suits which kind of work – Sonnet, Opus or something else – we cover in a separate tip.
You call another session with @
Messaging between Claude Code sessions passes native Windows by – it runs on macOS and on Linux, including Linux under WSL. Version 2.1.232 added the @ character to the input box: type it, pick the name of another Claude Code session, and write it a message. This extends the mechanism we covered a week ago, when sessions gained the ability to message each other – back then you had to know the command, now the name is enough. Sessions open on one machine also keep their names unique: a second session claiming a name in use gets a variant, and is told about it.
Two rows appeared in the /config panel for this. Messages from your other sessions decides what happens to a message arriving from another of your sessions, and Dialog expiry sets how long the question asking you to approve such a message waits for an answer.
Want more of Claude Code’s configuration secrets? Read our tip on the settings worth changing.
Claude went down on Sunday night
On Sunday 16 August at 21:58 UTC, logging in to claude.ai, Claude Code and Cowork stopped working, and errors on ordinary requests followed shortly after. Anthropic reported everything back to normal at 22:34 – the outage ran for over half an hour and took in the API and the console as well. The company hasn’t said what caused it.
There is one practical thing to take from this, and it isn’t about the outage itself. When Claude Code behaves strangely – no answer, a login error, a response cut off halfway – the first instinct is to look for the fault at your end: in the configuration, in a key, in the network. Before you fix anything, open status.claude.com. That is the page where Anthropic posts ongoing incidents, listed separately for claude.ai, Claude Code, Cowork and the API. Thirty seconds, and it saves you an evening spent repairing something that isn’t broken on your side.
Every service goes down sometimes, and half an hour is genuinely short. The interesting part is elsewhere: a year ago, the list of things you check in the morning before starting work ended at your mail and your calendar. It has one more item on it now, and nobody made that decision on purpose.