Events decouple producers from consumers, and that is genuinely valuable. What is less often said out loud is what you give up.

What gets harder

• Debugging. There is no call stack across a broker, so tracing has to be designed in from the first day rather than added after the first incident.
• Ordering. Anything that assumes two events arrive in the order they were emitted will eventually be wrong.
• Meaning. A published event becomes an API. Changing its shape is a breaking change to consumers you may not know about.

What makes it work

Version events explicitly and treat the schema as a contract with a registry behind it. Make every consumer idempotent, because redelivery is a promise the broker makes, not a failure. And publish facts about what happened rather than instructions about what to do, or you have built remote procedure calls with worse ergonomics.