see how the team improved Computer Use in @tryreplicas !
personally, this has improved my iteration speed significantly since workspaces can now fully setup, test, and run services on a machine as if it were any other human. I can also hop into the machine through the dashboard if I want to look around and inspect the running services.
more to come!
3 weeks ago, we launched Computer Use. Coding agents could test anything E2E on a Linux desktop, and provide screen recordings with every pull request.
In the beginning, agents failed simple tasks and sent us 20 min videos that no one watched.
To fix this, we went heads down and rolled out huge updates to the interface to increase agent accuracy and the UX of human reviewers, with a 20% gain on task completion.
(1/4) Read more π
small UI/UX update. @tryreplicas workspace sidebar has a new look. Now shows
- PR status icon
- most recent message (with icon from agent provider)
- compact options menu with more actions
- how long ago the workspace was doing work
- quick links to copy branch name and access GitHub PR/GitLab MR
Replicas is a very good yet stupidly simple idea.
Engineers often run claude code -p in CI to have agents review code.
This gives these agents a dashboard and observability, and lets you deploy from CLI/Slack/Linear.
A+