Astra came out last week and, while there is a lot to unpack in the release, the thing I’m most interested in is what it signals about the future of computer use.

For the last few years, most AI workflows have worked in one of two ways: either we bring the work to the model by copying something in or uploading a file, or we connect the model to other tools through integrations, APIs and automations so it can take action.

Both approaches have been incredibly useful, but they still rely on us creating a path between the AI and the work. Computer use starts to change how we interact with AI. 

Models are getting better at navigating the same interfaces we do, clicking through software, moving information between systems and figuring out the steps required to get from a goal to an outcome. This week, we are going to experiment with computer control.

Before we start: if you are getting this email forwarded to you and are not yet part of the lab - you can join the tiny AI lab here.

Find one task you would normally do manually on your computer (using AI) and see how much of it you can hand over to an AI agent.

You don’t need to pick anything complicated or particularly useful. The point is really to get a feel for what changes when the model can start operating the computer instead of just giving you an answer and if you have tried this in the past just how much models have improved. 

If you have permission to experiment with work systems, you can try something that moves between tools — maybe information from an email needs to end up somewhere else, or a task you normally do involves jumping between your browser, a spreadsheet and another piece of software.

Drop this into your AI of choice to get started:

Give it something you would normally spend an hour doing and let it work across a few different tools. This time, be explicit about the boundaries.

For example:

Here’s the outcome I need. You can research, navigate between tools, create files and make reversible changes without asking me. Ask for approval before you send anything, delete anything, spend money or make a change that would be difficult to undo.

Then see how far it gets. Be mindful of what model you are using so you don’t quickly run out of credits.

Astra and the marketing of the product are interesting because it’s not just the models getting better at performance. For the last few years, a lot of progress has been measured through better reasoning, better writing, better coding, longer context windows, fewer mistakes.

We’re obviously going to keep seeing those improvements, but computer use introduces another dimension of capability.

It’s one thing for a model to understand what needs to happen or give you step by step instructions. It’s another for it to actually navigate the computer and make it happen.

Until now, making AI useful inside an organization has often required us to build quite a lot around the model. We upload the information, create the integration, connect the API or build a workflow that tells different systems exactly how to interact. Computer use gives models another way in.

If the model can operate the software humans already use, it doesn’t necessarily need a purpose-built integration for every single task. It can increasingly navigate the interface, move between tools and figure out some of the steps on its own. That makes a much wider range of work potentially agentic.

Hit reply and tell me what you thought of the challenge this week.

I read every response, and your experience helps me shape these challenges. If you get stuck or have questions, reach me at [email protected]