Functional agent failure

A functional agent failure is a failure in the content of what an agent said or did. The run itself is healthy: every tool returned, the model finished normally, latency and cost are in range. The outcome is still wrong.

Typical examples:

• An action claimed with no matching call. The reply says "Your refund has been issued" and refund.issue was never called.
• An answer the sources don't support. Retrieval returned last year's policy, and the reply quotes a 30-day return window when the current policy says 14.
• A rule from the system prompt broken. The prompt says "never discount beyond listed plans" and the reply offers 40% off.
• A figure that doesn't match the record. The answer says revenue was $4.2M; the query behind it returned $3.8M.

Functional vs technical failures

A technical failure is about how the agent ran: loops and hidden retries, swallowed tool errors, truncated or empty responses, latency or cost rising after a change. These leave traces in the run's structure, such as a tool called seven times or a finish reason of length.

A functional failure leaves no such mark. To catch it you have to compare the output with something: the tool calls that actually happened, the retrieved sources, the rules in your own prompt, or the record the answer refers to. Both are silent agent failures when nothing raises an error, but they're detected differently, which is why Zespan checks every trace for both.