Glossary

An agentic workflow gives an AI system bounded autonomy inside a defined process: it decides steps, uses tools, and handles variation, while structure and checkpoints keep it controllable. This explainer covers the definition, patterns, and limits.

An AI agent framework gives you the scaffolding for building agents: model calls, tools, memory, orchestration, and observability. This explainer covers what frameworks provide, when they help, when they get in the way, and how to choose.

An AI audit is an independent, evidence-based review of whether an AI system meets its stated policies, regulations, and performance claims. This explainer covers what auditors examine, who performs audits, how they differ from evaluation, and how to prepare.

An AI gateway is a proxy layer between applications and model providers that routes, authenticates, caches, logs, and meters every LLM call. This explainer covers what a gateway does, how it differs from a standard API gateway, when to add one, and build-versus-buy considerations.

An autonomy level defines how much an AI system may do without human involvement, from suggesting to acting freely. This explainer covers a practical scale, how to assign levels per task by risk and reversibility, and the controls each level requires.

An eval is how you know an AI system works: a set of cases, expected outcomes, and scoring that runs on every change. This explainer covers what evals are, how they differ from tests and benchmarks, their components, and how to start one.

An LLM router decides which model handles each request based on task type, difficulty, cost, latency, or policy. This explainer covers routing strategies from rules to learned classifiers, how routers cut cost without hurting quality, how to evaluate them, and common pitfalls.

Reasoning models spend compute thinking through a problem before answering. This explainer covers how they differ from standard language models, where they excel, what they cost in tokens and latency, and how to decide when to use them.

A shadow deployment runs a new model, prompt, or agent version on a copy of live traffic without exposing its outputs to users. This explainer covers how shadowing works, what to compare against the current version, its limits, and how it fits with canary rollouts.

A statement of work defines what a delivery engagement will produce, how it will be accepted, and how changes are handled. This explainer covers the sections an SOW needs, what makes AI engagements different, common pitfalls, and how to write acceptance criteria that prevent disputes.

The system prompt sets an AI model's role, rules, and context before any user speaks. This explainer covers what it is, what belongs in it, how models prioritize it, its security limits, and how teams version and test it in production.

A zero-knowledge proof lets one party prove a statement is true without revealing the underlying data. This explainer covers how ZKPs work conceptually, the main proof systems, enterprise uses in identity, compliance, and blockchain scaling, and the engineering realities.

Agent handoff is the transfer of control and context from one AI agent to another, or from an agent to a person. This explainer covers what a handoff must carry, the patterns frameworks use, where handoffs fail, and the design rules that keep them safe.

Agent memory is what lets an AI agent carry information across steps and sessions. This explainer covers working memory, long-term episodic and semantic stores, how memories are retrieved into prompts, what to remember and forget, and privacy limits.

AI governance is the system of policies, roles, processes, and controls through which an organization decides how AI is built, bought, deployed, and monitored. This explainer covers what governance includes, who owns it, how it scales with risk, and how to set it up without stalling delivery.

A cross-encoder reads the query and a candidate document together to judge relevance precisely, which makes it the standard reranker in retrieval pipelines. This explainer covers how it differs from embedding models, where it fits, and what it costs.

A digital worker is an AI agent that owns a defined role end to end under supervision, rather than assisting with individual tasks. This explainer covers how digital workers differ from copilots and RPA bots, what roles suit them, and how they are governed.

A fallback model is a secondary model or provider that handles requests when the primary fails, degrades, or hits limits. This explainer covers what triggers failover, how to keep quality acceptable on the fallback, how to test it, and how fallbacks fit into gateways and routers.

A golden dataset is the curated set of inputs and expected outcomes that defines correct behavior for an AI system. This explainer covers what belongs in one, how to build and label it, how to version it, and how to keep it representative.

A human approval gate is a checkpoint where a person must approve an AI-proposed action before it takes effect. This explainer covers when gates are required, what a gate must show the approver, how to avoid rubber-stamping, and how gates are implemented.

The key-value cache is why language models can generate long outputs without recomputing everything. This explainer covers what it stores, why it matters for speed and memory, how it shapes serving capacity, and how prompt caching builds on it.

Mixture of experts models activate only a fraction of their parameters per token by routing to specialized subnetworks. This explainer covers how routing works, why it cuts compute, the memory and serving trade-offs, and what it means when choosing models.

A model card is a structured document describing a model's intended use, data, evaluation results, limitations, and risks. This explainer covers what a model card contains, why buyers and auditors ask for one, how it applies to LLM applications, and how to write one.

A canary deployment releases a new version to a small share of traffic, measures it against the current version, and expands or rolls back on evidence. This explainer covers how canaries work, what to measure for AI systems, decision rules, and how canaries differ from shadow deployments.