Agent reliability: is anyone getting past 80%?

I've been testing a few autonomous agent frameworks for internal workflows, and I keep hitting a ceiling around 80% task completion without human intervention. The last 20% is always edge cases—permission issues, malformed inputs, or the agent just going down a rabbit hole. Curious if anyone here has actually gotten reliable long-running agents into production. What's your stack, and how are you handling error recovery? Also, are you using structured prompts, strict tool schemas, or something like a human-in-the-loop fallback? I'm starting to think the real bottleneck isn't the model but the orchestration layer. Would love to hear what's working for you, especially around observability and logging agent decisions.