Strong opinions on multi-agent architectures
Originally posted to LinkedIn · October 2025
You can't enforce API contracts when your interface is natural language.
I have pretty strong opinions about multi-agent architectures, and finally took the time to write some down.
For as long as I’ve worked in conversational AI, I’ve been pitched the idea that “we’ll build an army of specialized agents and delegate to them on specific topics.”
It’s an obvious thing to try — microservices are popular for good reason. But they do not work for AI agents. It wasn’t a good idea 5 years ago, and it still isn’t now. You’ll end up with a distributed monolith.
And with all the recent protocol hype, the pitch has gotten even bolder: multiple agents collaborating unlock “emergent capabilities” that single agents cannot achieve.
That’s handwavy BS. The hard part isn’t communication but orchestration. And you can’t enforce API contracts when your interface is natural language.
There are better ways to scale AI agents to cover thousands of use cases: treat your sub-agents as dependencies, not separate services.