A computer for every agent
By Cursor
Summary
## Key takeaways - **Agents test like humans**: Cursor's agents use a computer to test changes and record videos of their work, just like a coworker would. This avoids the frustration of untested PRs. [00:10], [00:26] - **Onboards Excalidraw in minutes**: The agent successfully set up the environment for Excalidraw, a popular open source whiteboarding tool, and verified it works via video artifact. [00:22], [00:26] - **Adds table support in 40 minutes**: Pasting a GitHub issue link, the agent worked 40 minutes to add table support, navigating locally to insert the new table template successfully. [00:43], [00:57] - **CSV pasting via clever URL hack**: For CSV table requests, the agent opened a new tab to add data via URL, copy-pasted it into Excalidraw, and handled edge cases like quoted commas. [01:22], [01:49] - **LaTeX math rendering demoed**: The agent added LaTeX support, demoing E=MC², quadratic formula, and integrals properly converted in the software. [02:19], [02:32] - **Take control for low-latency testing**: Users can click 'take control' for full desktop access in low latency to further test the agent's environment. [02:03], [02:10]
Topics Covered
- Agents Test Changes Like Humans
- Agents Build Table Support End-to-End
- Follow-Up Prompts Unlock CSV Parsing
- Agents Self-Fix CSV Edge Cases
- LaTeX Rendering Proves Math Mastery
Full Transcript
If your coworker asks you to review a PR and then you find out they didn't even test their changes, you're probably going to be a little upset. This is how it can feel working with coding agents.
Cursor's agents can now use a computer to test their changes and record videos of their work. Let me show you how. I'm
going to have Cursor on board to my codebase, which is Excaladraw, a popular open source whiteboarding tool. And we
can see the agent was able to successfully set up our environment. and
we get back a video artifact showing that the agent can use a computer to navigate around and use our software to verify that it works. So now that our environment is set up correctly, I went
through the GitHub repo and I found a few bugs I wanted to fix and a few features that I wanted to add. One of
them I thought was interesting was adding support for a table. So I pasted in a link to the GitHub issue and the agent worked diligently for about 40 minutes and it gave me back an artifact
showing that it could use a browser. It
could navigate to the software running on the correct port locally and successfully insert in this new table template that was created. So it's
pretty nice to see that this works as expected. But I wanted to take it
expected. But I wanted to take it further. So, I was looking through the
further. So, I was looking through the GitHub issues and I noticed there was also a request asking for pasting CSVs as table templates. So, I asked a follow-up question to the agent. It
worked again for another 20 minutes and I got back an artifact showing that now it could also paste in CSV data and give me tables. And the way the agent did
me tables. And the way the agent did this was uh very clever. It used a new tab to add data directly into the browser through a URL so that it could
copy paste it and then paste it in in a new tab. And I also thought it was
new tab. And I also thought it was interesting that the agent was trying to think through and fix edge cases itself.
So it came up with a CSV set where it wanted to use a comma which needed to be surrounded by quotes and it copy pasted that text inserted it into Excalibraw
and it was properly parsed. So that
worked pretty well. If I wanted to do further testing myself, I have full access to the desktop inside of this environment. So I can click take control
environment. So I can click take control and I get access to the full computer where I can in very low latency interact with this environment. Now the other
feature that I wanted to add was support for latte rendering for all of our math fans out there. And after this finished successfully, I got back a artifact
showing all of the classics. It's
demoing E= MC^2, the quadratic formula, even some integrals, it looks like. So,
it's able to successfully prove that it can type in math equations and then the math equations are properly converted into latte formatting. So, this works
pretty well. If we're happy with these
pretty well. If we're happy with these changes inside of the cursor agent, we can view all of the diffs and then when we're ready, we can have the PR, which is automatically opened, marked as
ready, and then review directly from inside Cursor. Everything that I just
inside Cursor. Everything that I just showed is now live inside Cursor. And if
you want to try it out, you can go to cursor.com/onboard
cursor.com/onboard to set up your repo. We'd love to hear your feedback.
Loading video analysis...