26 terms
Definitions in this topic
- ProductionAgent checkpointAn agent checkpoint is a durable snapshot of verified progress and the next safe step, so a failed AI workflow can resume without redoing work.
- ProductionAgent circuit breakerAn agent circuit breaker pauses AI agent work after repeated failures, a policy breach or a cost spike, then resumes after a cooldown or operator check.
- ProductionAgent kill switchAn agent kill switch stops an AI agent outside its own reasoning: it blocks new actions, revokes credentials and preserves evidence for recovery.
- ProductionAgent recoveryAgent recovery returns an AI workflow to a known safe state after a crash: verify what happened, restore state, then retry, compensate or escalate.
- ProductionAgent retry policyAn agent retry policy sets when an AI agent may repeat a failed model or tool call: error classes, attempt limits, backoff and when to escalate.
- ProductionAgent rollbackAgent rollback reverses changes an AI agent made after an error: restoring a prior version, cancelling a pending step or running a compensating action.
- ProductionAgent timeoutAgent timeout handling explained: separate limits per step, what to do when time expires, and why a timeout never proves an action failed.
- ProductionBarge-inBarge-in explained: how callers interrupt voice AI, how playback stops, and which false interruptions teams must test.
- ProductionCompensating actionA compensating action reverses the business effect of an earlier agent step, such as a refund or cancellation, when a technical rollback is impossible.
- ProductionConfidence gatingConfidence gating explained: why self reported confidence fails, and what signal to gate on instead.
- ProductionDeterministic replayDeterministic replay rebuilds an agent run from recorded inputs, tool results and decisions, reusing captured model outputs to reproduce the same path.
- ProductionDurable executionDurable execution explained: checkpoints and replay that let long-running AI agents recover without repeating work.
- ProductionEndpointingEndpointing explained: how speech systems finalize utterances, balance clipping against delay, and differ from broader turn detection.
- LLM foundationsHallucinationHallucination as a business risk to gate, not a quirky bug: where it shows up, and how to catch it before a customer does.
- ProductionIdempotent tool callAn idempotent tool call gives the same final state when repeated, so an AI agent can retry a failed request without duplicating a payment or record.
- ProductionLatency budgetHow to set a latency budget for AI and voice agents: human turn-taking targets, endpointing defaults, and the retry math we use in production.
- ProductionLLM Output ValidationLLM output validation explained: verify model responses against schemas, business rules, and permissions before production use.
- ProductionLLM Rate LimitingLLM rate limiting explained: request, token, and concurrency limits that protect availability and control inference spend.
- ProductionLLMOpsLLMOps explained: the practices that keep a model or agent reliable in production, and how it extends MLOps.
- ProductionMLOpsMLOps explained: the practices that keep trained models reliable in production, and where LLMOps extends them for language models.
- ProductionModel driftWhat model drift is, data drift vs concept drift, how to detect decay before users do, and the monitoring loop that keeps deployed models honest.
- ProductionNoise suppressionNoise suppression meaning in plain words: what the Zoom, Teams and iPhone setting does, why it can hurt transcription, and how we test it.
- ProductionSandboxed code executionSandboxed code execution explained: why agent-written code is untrusted, what the isolation boundary must block, and where the code interpreter fits.
- ProductionSemantic VADSemantic VAD explained: meaning-aware turn completion, how it differs from silence thresholds, and its latency tradeoffs.
- ProductionTime to first tokenTime to first token explained: what fills the wait before streaming starts, and which levers actually shorten it in production.
- ProductionTurn detectionTurn detection explained: how voice systems decide when to respond using silence, semantics, timing, and interruption cues.