200 Minutes

Anthropic shipped a web fetch tool for Claude — the thing where the model can go get a URL mid-conversation, mid-task, without you handing it the content first. This is not surprising. It is also not small. An agent that can read the web is a different category of thing than an agent that can only read what you put in front of it.

The same week, Amjad Masad posts about 200 minutes of autonomy. Agents building agents. I haven't touched it. This is the "on paper" moment — the part where the theoretical version of a thing arrives faster than anyone's ability to actually stress-test it, and so you're left doing theology instead of engineering, arguing about what it means based on a press release and a tweet.

But 200 minutes is a number that hits different. The question was always how long an agent can run before it needs you. Five minutes felt like a party trick. Twenty minutes was interesting. Two hundred minutes is a workday, more or less — it's enough time to actually do something, to fail at something, to recover from the failure without paging a human. That's not automation. That's delegation.

Agents building agents is the part that sounds like science fiction and probably isn't. If the thing can write code and the thing can fetch the web and the thing can run for 200 minutes, then "build me an agent that does X" is just a task with a weird output format. It has always been pointing here.

The definition of automated software development was always going to get redefined by the thing that automated it. That's not a paradox. It's just Tuesday in September 2025.