Aller au contenu
Implementa.
← Full glossary

Semantic anomaly

reliability

semantic failure · meaning-level failure

A failure where every agent call returns success and latency looks normal, yet the outcome is wrong: it took the wrong branch, misclassified, or executed valid actions in an unsafe order.

A semantic anomaly is the failure class that conventional logs and traces cannot see, because everything observable is green. A tool call can succeed technically and still be wrong in meaning: cancelling an appointment before confirming its replacement, filing an expense under the wrong category, invoking valid APIs in an order that cannot be undone. Recent academic work on agent system operations puts this at the centre: it treats observability as a semantic problem rather than a system-health problem, and separates intra-agent anomalies (reasoning, action, memory, security) from inter-agent ones (task specification, orchestration, communication, termination). The practical consequence for whoever operates the system: functional correctness and operational safety are two different measurements, and an agent can end up leaving the right final state after doing something irreversible on the way.

How it differs from

Technical failure
A technical failure throws an error, a timeout or a 500 and surfaces by itself; a semantic anomaly returns success and has to be looked for.
Agent observability
Observability is the ability to look; the semantic anomaly is the failure class that ability misses when it only measures system health.

Examples

  • The billing agent issues all 240 invoices for the month without a single API error and books 18 of them to the wrong fiscal year.
  • The support agent closes the ticket with a well-written, correct answer to a question the customer never asked.

FAQ

How do you detect it when the logs are green?
By comparing the output against expected meaning rather than against system state: evals on cases with known answers, business rules applied to the output, and human sampling of closed cases. That is evaluation work, not monitoring work.

Related terms

Sources & further reading