The best thing i added to my agent was permission to say "not sure"
retained text
For months every step in my agent had to answer yes or no, pick a tool, pick a category so it always did, even when the input was garbage. The bad calls weren't the hard cases tho they were the weird ones, a half filled form, a ticket in two languages, a screenshot where text should be anything man, the model would just pick something with full confidence and move on. I gave every decision step a third option as such, not sure, here's why anything under a threshold goes to a human queue or a bigger model instead of guessing. Now about [X]% of calls land in "not sure", the wrong answer rate on the rest dropped to [Y] and the human queue is small enough that someone actually reads it. The catch is the threshold only works if the confidence numbers mean something, plenty of models will say 0.95 on stuff they're making up and some decision models are built around this, I think it was [bivia](https://articul8.ai/bivia/) that has abstain as an actual output mode but whatever you use, check that 0.9 really means right 9 times out of 10 before you trust it. Anyone else running agents with an explicit path?
document fetched 2026-11-01 00:27 · canonical https://reddit.com/r/SaaS/comments/1xk6605/the_best_thing_i_added_to_my_agent_was_permission