In a conversation about how the team behind Claude Code works themselves, one offhand line stuck with me. The question was simple: when you sit and work alongside the agent, what is your go-to feature? The answer from Boris, one of the makers, was simple too. “That used to be Plan mode. I do not use it anymore.” And instead? “Auto mode. For everything.”
It made me laugh, because I caught myself in the exact same shift. Without ever deciding to, Plan mode had quietly left my workday. This piece is about why, and about the one moment where I still hesitate.
What Plan mode was for
Back up to what it did. In Plan mode the model first works out an approach and lays it before you, before a single line of code changes. You read the plan, approve it, and only then does it get to work. An in-between step that builds trust.
That step existed for a reason. With the older models, roughly Opus 4 through 4.5, that planning moment genuinely helped. The model benefited from an explicit round of thinking before it began. But Boris points at something I notice too: from the newer models on, certainly from 4.7, that planning step is no longer needed. The model plans along the way, while it works. The separate round up front adds little now.
Why Auto mode fits better
Auto mode flips it around. You give your instruction, the agent starts straight away, and you move on to the next thing. No plan sitting there waiting for your sign-off. For me that is the whole gain: I start an agent, and instead of staying parked next to it watching, I start the next one. Or I go get coffee.
And what about safety, you think right away. The old way gave you a permission prompt on every sensitive action: may this command run, yes or no. Sounds careful. But something human happens: if you click “yes” nine times out of ten, your eyes slide right over the tenth. Auto mode has a separate model watching along that blocks suspicious actions. Not you waving through a hundred routine requests, but attention that goes only to the one thing that actually matters. That is, against my first instinct, safer than reading everything yourself.
And yet, when brainstorming
Here is the honest part. Sometimes, when I am still thinking something through rather than building it, Plan mode still feels nicer. As if I pull up a chair and we think out loud together before anything happens.
But when I test it, that feeling does not hold. Auto mode arrives at just as good a plan, it just writes around it right away. The outcome is no worse. So it is a fair question why I still switch Plan mode on now and then.
The answer is not in the tool, but in me. I simply like that plan on my screen. It gives me a sense of grip, even when it adds nothing functionally. Boris says exactly that in the same conversation: some people keep using it because they like having that in-between document. So it is not a necessity. It is a habit that reassures.
And there is nothing wrong with that, as long as I know that is what it is. A preference is not the same as a requirement. It is worth noticing when you do something because you must, and when because it feels familiar.
The proof
The nice thing is that the evidence sits right under this piece. This series, and most of the articles on this site, I make with Auto mode. The agent runs, I think along, it works. No plan asking for my signature first, and still not a moment where I lost the grip.
That is the same line running through my other pieces about AI that executes: let it do the work, as long as I see what happens and choose myself when it goes to full power. Letting go of Plan mode is a small exercise in that. Not because the in-between step is bad, but because I no longer need it, and wanted to feel that for myself before I dropped it.
This is the first part of a series on how the makers of Claude Code work with it themselves, and what I take from that. More on AI as an operator: how I left WordPress and one word, a fleet of agents.