Whenever a new foundation model provider announces some big update, it’s become custom to check in on your early stage founder friends and see if any of the R&D they’ve spent in the last few months has just become sunk cost.

It happened with o1, and again with Anthropic’s 10/22 release for Sonnet 3.5–particularly when paired with the computer use demo.

In our case, the biggest takeaway wasn’t computer use itself (a really cool demo of something these models have been able to do for months now), but what it signals about where Anthropic (and OpenAI) are going.

It’s pretty clear the bulk of their model improvements are focused on moving away from chat-based back-and-forth workflows, and instead enabling long-running tasks with minimal human intervention.

The buzz word you’ll hear McKinsey and friends throwing around is “agentic” a word that up until now has been almost entirely unremarkable, but you’ll soon be sick of hearing about.

For now it’s great for us: the models are better at the types of things we need them to do. As the user-friendly products like Claude, ChatGPT, Gemini, etc. integrate these as first-class features, I think a lot of companies will need to ask themselves how much of their “agent infrastructure” is worth against a $30 a month subscription.