Anthropic's AI Now Writes 65% of Its Own Team's Code

Anthropic built an AI coworker that lives inside Slack, and it's already writing 65% of their product team's code.

They call it Claude Tag. You @ it like you'd @ any teammate who actually reads the docs. It doesn't reply with a suggestion. It goes and does the work.

Here's the part that matters more than the percentage.

One Claude Per Channel Changes the Math

There's one instance per channel, not one per person. That means everyone in the channel sees what it's working on. Nobody has to ask "wait, did someone already handle this" because it's sitting right there in the thread.

It also remembers the channel. You stop re-explaining the same context five times to five different tools or five different threads. You tell it once, and it stays told.

And it works async. You hand it a task and go to lunch. You come back to finished work instead of a status update asking for clarification. That's the actual unlock. Not that it's fast, that it doesn't need you standing over it.

Your Team Just Got a Coworker Who Doesn't Sleep

Think about what 65% of a codebase means coming from something that never asks for a raise, never takes PTO, and never gets passive-aggressive in a code review comment.

Your most reliable coworker now runs on tokens and never takes a day off.

This isn't a demo or a lab experiment. This is Anthropic's own product team, in production, right now. The tool writing the product is embedded in the same Slack channel where the humans plan the product.

If your team is still treating AI like an autocomplete you copy-paste from, you're behind a company that's letting it @ mention its way through actual sprint work. Go look at how your team hands off tasks and ask honestly whether a channel-aware AI coworker would slow you down or just expose how much of your process was busywork all along.