Definition
A failure that looks mechanically successful while producing the wrong meaning, intent, binding, or real-world consequence. The process may return a valid response, call an available tool, pass a syntactic check, or exit with status zero. What it actually did does not satisfy what the user or system intended.
Examples include updating the wrong customer's record, summarizing a source that was never retrieved, applying a correct policy to the wrong jurisdiction, or declaring a deployment healthy because the command succeeded while the user-visible behavior remains broken. These failures are dangerous because ordinary availability and error metrics can record success.
Why agents are vulnerable
Tool-using agents carry meaning across multiple steps: a name becomes an identifier, an instruction becomes a plan, a plan becomes tool arguments, and a tool result becomes a claim. Binding drift or an unsupported assumption at any step can survive the rest of a mechanically valid trajectory. The final narrative may even explain the wrong action convincingly.
Distinguish it from nearby terms
A syntax error violates form. An operational failure interrupts execution. A semantic failure completes the operation with the wrong meaning or consequence. A silent failure is any failure not surfaced to the responsible observer; semantic failures are often silent, but the terms are not exact synonyms.
Check your understanding
The tool call returned 200 and the agent says the invoice was corrected. What must you verify? The invoice identity, field binding, before-and-after state, applicable rule, and downstream customer outcome. Transport success alone cannot establish semantic success.