Update OSWorld blog post to correct link formatting for the dokur project series, enhancing clarity and consistency in references.

Author: Timothyxxx

data/blog_posts/osworld-verified.md

@@ -114,7 +114,7 @@ Computer use agent development environments present unique complexities (after a
 
 We have been continuously searching for the technical optimal solution. Our initial attempt involved building controlled environments using VMware, where we distributed `.vmdk` files and `.vmx` configuration files, saving images during first-time execution on users' computers. However, this approach had significant drawbacks: despite being theoretically parallelizable, the software consumed substantial personal computer resources with excessive runtime. Additionally, VMware became increasingly closed and cumbersome to download after Broadcom's acquisition.
 
-Inspired by the dokur project series (https://github.com/dockur/windows), we integrated Docker containerization technology 2024 Summer, utilizing an open-source VMware-like service - QEMU - running within Docker containers to execute virtual machines. This approach enabled us to achieve multi-environment parallelization on a single server, running 8 or even 16 environments simultaneously for experiments, though still constrained by server performance. We also implemented AWS support during this period but didn't pursue it extensively.
+Inspired by the [dokur project series](https://github.com/dockur/windows), we integrated Docker containerization technology 2024 Summer, utilizing an open-source VMware-like service - QEMU - running within Docker containers to execute virtual machines. This approach enabled us to achieve multi-environment parallelization on a single server, running 8 or even 16 environments simultaneously for experiments, though still constrained by server performance. We also implemented AWS support during this period but didn't pursue it extensively.
 
 Later, WindowsAgentArena [[8]](#ref8) (whose leading authors founded the c/ua company) left a profound impression on us by leveraging cloud services for parallelization, compressing evaluation time from 10+ hours to just 20 minutes. We think it is the right direction, while enhancement can always be done. So we actually followed WindowsAgentArena's approach by leveraging AWS as cloud services and extended this feature in OSWorld infrastructure, which enables us to run up to 50 environments simultaneously and shorten evaluation time to minutes while ensuring comparability across evaluations.