Why do AI agents still fail at simple multi-step tasks?

I've been testing several agent frameworks lately, and honestly, the hype doesn't match reality. Give them a straightforward workflow—like fetching data from an API, transforming it, and posting it somewhere—and they'll often get stuck on the second step. Tool calling works fine in isolation but falls apart when context gets long or the API returns something slightly unexpected. I'm curious what others are seeing. Are you relying on agents in production, or just for prototyping? And for those who've made them work reliably, what's the secret? Strict schemas, smaller subtasks, or just avoiding the fancy stuff entirely? I'm close to going back to plain functions with LLM calls sprinkled in.