Inside Anthropic’s 30,000-agent workforce

Anthropic is measuring how much AI builds AI

Claude now “leads” 26% of Anthropic’s AI R&D work, up from under 1% in February.

That means it can complete most of those tasks from a high-level prompt while a human supervises. More than 90% of Anthropic’s measured R&D already involves Claude at the “collaborates” level or higher.

The important part is the speed. Six months is a short time for AI to move from assisting researchers to handling meaningful chunks of the work that produces future models.

Anthropic also says roughly 30,000 agents are working simultaneously on research and engineering inside its most-used internal platform. Their actions pass through automated monitors, with about one in 47,000 decisions blocked during August.

You can picture the human engineer reading the final write-up before deciding whether a fix ships that night.

The open question is how long that human checkpoint remains necessary.

Google is giving the household its own AI agent

Google’s new family agent gets its own Google account.

CC can work with up to six household members, pulling in whatever emails, calendars, Drive files and messages each person chooses to share. Every morning, it can turn that information into a shared brief covering appointments, tasks and changes to the day.

This feels more consequential than another chatbot feature because Google is putting an agent directly inside the messy administrative work families already do. Think school emails arriving after dinner, a photo of a soccer schedule, or a permission slip sitting in a PDF.

CC can update calendars, fill forms, make shopping lists and check live drive times between activities.

The interesting part is its shared memory. Google says CC can distinguish household information, like favorite restaurants, from details belonging to individual members.

If families actually trust it with this access, the household could become one of the clearest tests of whether persistent AI agents are genuinely useful.

Open models need safety that survives modification

Hugging Face already hosts more than 6,000 models whose safeguards have been deliberately removed.

Baseten’s Base Labs is now working with Hugging Face and Goodfire to develop safety evaluation and monitoring tools for open-weight models. The group plans to publish its methods and bring other developers into the effort.

I think the 6,000 figure explains why this matters. Once model weights can sit on a developer’s laptop and be modified freely, safety measures added before release may have a short shelf life.

The partnership is trying to move some protection closer to training and deployment, where developers serving these models could continuously monitor their behavior. Goodfire’s work on model internals could be especially relevant here, although the companies haven’t explained the technical system yet.

That missing detail matters.

If the resulting tools stay useful after models are modified, open-weight AI could get a more practical safety model than relying on safeguards shipped with the original release.

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.