Knowledge boundary detection is an agent’s ability to recognize when a question or task lies outside the knowledge it can reliably use. Instead of producing a plausible guess, the agent signals uncertainty or shifts into graceful degradation mode. section 10.1.1 section 10.3.6
This differs from simply generating an answer. Metacognition asks whether the agent knows enough to answer, especially when retrieval is weak, evidence conflicts, or the correct answer is absent. section 10.1.4 section 10.1.1
A practical boundary check combines several signals:
- Confidence state: include the agent’s internal confidence rather than relying on verbal claims such as “I’m confident.” section 10.1.4
- Boundary response: classify the situation as within knowledge, at the edge, or outside knowledge, then gate execution or degrade gracefully. section 10.3.6
From confident guessing to a boundary-aware response
An agent asked about a topic absent from its retrieval and memory coverage should not present a plausible answer as fact. It should recognize the boundary, report limited knowledge, and avoid silently eroding user trust. A practical implementation averages retrieval relevance, memory coverage, and confidence to select a boundary state. section 10.1.1
