Steps to reproduce:

1. Have a hobby like What's up, home? and get this brilliant idea of setting up an ugly kludge with in total of six Zabbix agents running on a single machine for making it possible to perform blog global availability checks over Proton VPN.
2. Setup virtual Ethernet NICs on your Raspberry Pi 4 and Raspberry OS 12
3. Point each Zabbix Agent instance to point on its own IP address
4. Setup new hosts on Zabbix, each pointing to these dedicated IP addresses
5. Create a new template on Zabbix with just a single web scenario in it, the scenario being "load https://whatsuphome.fi/ and see how that goes"
6. Apply the template for the previously created hosts
7. See how response times around the globe are similar
8. Go back to your template, add another item, this time add web.page.perf[https://whatsuphome.fi/]
9. See how response times around the globe are like they should 
10. For more, read https://whatsuphome.fi/whatsuphome/part68 (blog post evolving whilst I type this bug report)

Result:

With traffic going to five different VPN tunnels pointed across the globe, I would expect the response times to vary heavily. Instead, when using web scenarios, they are pretty much the same. With web.page.perf item, everything works as should. So, I suspect the web scenarios are using the primary NIC of the machine instead of the IP that is assigned for the agent.
**

Expected:

Web scenarios should obey the agent interface that is defined for the host.