Term Definition
Silent Failure
A failure mode where an AI agent stops executing correctly or deviates from intended logic without throwing an error or alerting the system.
Full Explanation
Silent failures are the most dangerous form of agent failure. Unlike a hard crash (which triggers a 500 error or process exit), a silent failure occurs when the agent believes it has completed a task successfully but the external state (CRM, Database, Git) remains unchanged. This often happens due to unhandled tool-call exceptions or 'hallucinated' confirmations.
Why it Matters
Standard monitoring tools like Sentry or Datadog won't catch silent failures because no exception is raised. StateBase detects these by verifying state transitions against the agent's reasoning trace.