Confidence gating is a design pattern where an agent acts on its own only above a measured confidence threshold, computed from a signal you can actually verify, and routes anything below it to a person, instead of trusting the model's own claim about how sure it is.
Why does self reported confidence fail as a gate?
A model asked to rate its own confidence learns to sound sure, not to be right; the number drifts toward the top of the scale whether the answer deserves it or not, a failure mode covered in why self reported confidence is not a reliable signal. Gating on that number manufactures trust exactly where it has not been earned.
How does confidence gating actually work in an agent?
The agent computes a real signal, a retrieval score, a validator result, agreement across repeated runs, and a plain deterministic rule compares it to a threshold. Above the line, the AI agent proceeds; below it, the case routes to a human in the loop instead of a blind retry. This is Soba’s own operating pattern: guardrails catch policy violations at runtime, while confidence gating decides which individual cases need a person’s judgment, and the threshold itself gets tuned against what LLM observability shows actually happened once the gate ran on real traffic.
Frequently asked questions
Why not just ask the model how confident it is?
Because that number is unreliable. A model's self reported confidence is shaped by training that rewards sounding helpful and sure, so it clusters near certainty whether the answer is right or wrong. Gating on it means gating on noise, not signal.
What should a confidence gate actually measure instead?
Whatever genuinely correlates with correctness in your domain: a retrieval score behind a RAG answer, how cleanly a rule matched, agreement across several sampled runs, a schema validation result, or a cross check against a system of record. All of those can be logged, plotted, and calibrated.