Hallucination refers to a model's fluent and confident output that is factually incorrect. It arises because the model generates text that is statistically plausible given its training data, rather than retrieving verified information. This behavior is an intrinsic property of how language models operate, not a defect to be patched. In an agent, a hallucination is not just a wrong sentence; it can become an action.
In an agentic system, a hallucination extends beyond a mere false statement; it can become a basis for incorrect actions. An agent might confidently invent a non-existent API call, misinterpret a system state, or fabricate a file path, leading to system errors, data corruption, or failed tasks. Mitigation requires external checks, such as validating outputs against known facts or using a second, independent source.
The common misconception is viewing hallucination as a problem that a stronger, more accurate model will eventually eliminate. Instead, it is a fundamental characteristic of a system designed for probabilistic generation of plausible text, not for deterministic fact retrieval. Relying on an agent's internal reasoning to detect its own hallucinations is unreliable; effective mitigation always involves grounding its outputs against external, verifiable sources or employing a harness that validates actions before they are executed.