For multi-agent workflows, how do you handle disagreements between agents?

Single agent in production is solved. Multi-agent introduces a problem nobody fully has a clean answer for.

When two agents reach different conclusions about the same task:

- Do you have a supervisor agent break the tie?

- Default to one agent's authority based on hierarchy?

- Escalate to a human reviewer?

- Average or merge outputs?

What's working in your stack? Or honestly, is everyone still in the 'we'll figure it out when it happens' phase?