agent-to-agent resiliency and observability

[salmanap]

What would you like to see with respect to resiliency and observability when agents are communicating with other agents. Arch uses the Envoy substrate so it can easily extend to this (agent mesh) scenario by moving crufty resiliency and observability work outside business logic. But curious to get thoughts from the community on what exactly you would like to see, what you think will be valuable. Of course machine learning related asks are very relevant as arch is designed with small LLMs