Been digging deeper into how LLM agents behave in real systems, and one pattern keeps coming up.
Things start to break when agents are asked to do work that requires exactness. What should be deterministic turns into something that only works most of the time — and that’s where reliability quietly falls apart.
Wrote a short breakdown with a real example and a simple way to think about the boundary:
→ https://grabduck.substack.com/p/where-to-draw-the-line-in-agent-systems