I think we're repeating the early microservices mistake with AI agents

A lot of agent demos remind me of what happened when microservices first became popular.

Everyone was excited about splitting systems into smaller components. It looked elegant in diagrams. It looked scalable. It looked like the future. Then people realized the hard part wasn't building services.

It was communication, orchestration, observability, debugging, versioning, and managing complexity.

When I look at multi-agent systems today, I get a similar feeling. Building an agent isn't particularly hard anymore.

Building 5, 10, or 20 agents that can reliably work together, maintain context, recover from failures, and remain manageable over time feels like a much bigger challenge.

Sometimes I wonder whether the next breakthrough in agent systems won't come from better models at all. It'll come from better engineering practices around agents.

Curious whether people building production systems agree or if I'm completely off here.