Uncertainty Markers in AI Agent Outputs - Why Knowing What the Model Doesn't Know Matters
Uncertainty Markers in AI Agent Outputs
One of the biggest trust problems with AI agents is that they present everything with the same level of confidence. Whether the model is certain about a fact or guessing based on pattern matching, the output reads the same way. Uncertainty markers change that.
The Confidence Problem
When an LLM generates a response, it has no built-in way to signal "I am not sure about this part." The model produces tokens with probability distributions, but those probabilities are not exposed to the user or the consuming system. A hallucinated fact looks identical to a verified one.
For chatbots, this is annoying. For AI agents that take actions based on their outputs, it is dangerous. An agent that confidently executes a wrong action is worse than one that pauses and asks for clarification.
How Uncertainty Markers Work
The WCY (What Can't You) approach asks models to explicitly mark sections of their output where confidence is low. Instead of generating a clean answer, the model produces structured output with confidence annotations. "The user's timezone is PST [high confidence] and their preferred language is French [low confidence - inferred from name]."
This gives downstream systems the information they need to decide whether to proceed automatically or escalate to a human. High-confidence outputs get auto-approved. Low-confidence outputs trigger a review step.
Practical Implementation for Agents
In an agent workflow, uncertainty markers become decision points. When the agent needs to click a button but is uncertain which button matches the user's intent, it should flag that uncertainty rather than guessing. When it generates a query but is unsure about a join condition, it should mark that ambiguity.
The pattern is simple: require structured output from the LLM that includes confidence levels, then build branching logic in the agent that routes low-confidence decisions to human review. The agent becomes more trustworthy not by being more accurate, but by being honest about its limitations.
Fazm is an open source macOS AI agent. Open source on GitHub.