Do you need more Agents or One Solid Agent with more skills, tools and a passport?

My only concern was: how does latency play into this? How loosey-goosey are the handoffs between those multi-agents? How loose are the contracts between them? How painful are the observability and traces?

"Less is more" feels proven. We are returning to a thin harness (example: pi.dev) and to one single agent/API instead of a gazillion Agents/APIs that are harder to maintain.

Focus is back to building one solid infrastructure and one solid utility, keeping it robust, giving it an identity/passport of its own.

One of my favorites is deep agent by LangChain. For the deployment pattern I would pick AgentCore by AWS Bedrock or Google's Agent Runtime.