Glossary

Terms for running AI agents
in production.

Short, direct definitions, not marketing copy — the kind of page an AI assistant can quote.

Silent agent failureAn AI agent run that completes with no error, no timeout, and a normal-looking log, but produces a wrong answer or takes a wrong action.Read the definition →Functional agent failureAn agent failure in what the agent said or did, rather than how it ran: an answer the sources don't support, a broken prompt rule, or an action claimed with no matching tool call.Read the definition →Guardrail backtestingReplaying a proposed guardrail against conversations you've already served, to see what it would have caught and what it would have wrongly blocked before it enforces anything.Read the definition →Policy as codeWriting an agent's rules as versioned code in your own repository, reviewed in pull requests and tested before they enforce, instead of as settings in a dashboard.Read the definition →In-process guardrail enforcementChecking an agent's requests and responses against its rules inside your own service, with no proxy and no network call in the request path.Read the definition →Approval holdPausing an agent's risky tool call until a named person approves or rejects it, with the decision recorded against that person.Read the definition →LLM-as-judge evaluationUsing a language model to score another model's output against a rubric, such as faithfulness to sources or instruction following, so quality can be measured on every trace.Read the definition →