Agentic AI Glossary: 30 Terms Executives Actually Need

This glossary defines the 30 terms that come up most often when executives evaluate agentic AI — each in 40-60 words, plain-English, with a note on why it actually matters to a buying decision rather than an engineering one.

Every other pillar article on this blog links its jargon back here; this page links back out to all of them, plus the services each concept connects to.

Core concepts

Agent

Software that pursues a stated goal by reasoning over its current situation and choosing among available actions, rather than executing one fixed script. Why it matters: this adaptability is the entire value proposition over traditional automation — see our full explainer.

Agentic automation

The practice of using AI agents — not fixed scripts — to automate business processes that involve judgment, exceptions, or variation. Why it matters: it's the category this entire site is about; see our agentic automation service.

Reasoning loop

The cycle an agent runs: evaluate the current state against its goal, decide the next action, take it, re-evaluate. Why it matters: this loop is what lets an agent recover from an unexpected input instead of just failing.

Tool use / function calling

An agent's ability to call external functions or APIs — look up a record, send an email — as part of its reasoning loop. Why it matters: an agent without tool use can only talk, not act; tool use is what makes it operationally useful.

RAG (retrieval-augmented generation)

A technique where a model retrieves relevant documents or data before generating an answer, grounding its response in real content instead of memory alone. Why it matters: RAG is a primary mitigation for hallucination in any agent that needs to cite facts.

Fine-tuning

Further training a pre-trained model on a specific dataset to specialize its behavior for a narrower task. Why it matters: most mid-market use cases don't need it — prompting and tool use solve most problems more cheaply; fine-tuning is for the exceptions.

Risk and control

Hallucination

When a model states something false with apparent confidence, unsupported by real data. Why it matters: it's the single most cited risk objection to agentic AI, and it's addressed by grounding claims in real tool calls rather than model memory — see our governance guide.

Guardrails

Explicit, engineered boundaries on what an agent may do without human approval. Why it matters: guardrails are what makes an agent's "confident wrong action" failure mode containable instead of catastrophic — see RPA vs Agentic AI on why this failure mode differs from traditional automation.

Human-in-the-loop (HITL)

A design pattern where a human reviews or approves an agent's action before, or instead of, full autonomy. Why it matters: it's the practical mechanism behind tiered oversight — most production agents should start here, not at full autonomy.

Evals

Structured tests that measure an agent's or model's performance against defined success criteria, run before and after changes. Why it matters: without evals, you can't tell whether a prompt or model change made things better or worse until a customer notices.

Observability

The practice of instrumenting a system so its internal behavior (reasoning traces, tool calls, outcomes) is visible after the fact. Why it matters: it's the difference between "the agent did something wrong" and "here's exactly why, and here's the fix."

Drift

Gradual degradation in an agent's or model's performance over time, often from changes in input patterns the system wasn't built for. Why it matters: drift is why "we tested it once" isn't governance — ongoing monitoring is required.

Red-teaming

Deliberately probing an agent for failure modes — prompt injection, unsafe actions, biased outputs — before it reaches production. Why it matters: finding a failure mode in testing is a non-event; finding it in production is an incident.

Prompt injection

An attack where malicious input tries to override an agent's instructions (e.g., a document the agent reads contains hidden text telling it to ignore its guardrails). Why it matters: any agent that reads untrusted content (emails, documents, web pages) needs defenses against this specifically.

Architecture and orchestration

Orchestration

Coordinating multiple steps, tools, or agents toward a larger goal. Why it matters: most real business processes need orchestration across systems, not a single isolated action.

MCP (Model Context Protocol)

An open standard for connecting AI models to external tools and data sources in a consistent way, reducing the custom integration work each new tool would otherwise require. Why it matters: standards like this lower the cost of connecting agents to your existing systems.

Multi-agent systems

Architectures where multiple specialized agents collaborate — one plans, one executes, one reviews — rather than one agent handling everything. Why it matters: complex, multi-system reasoning workflows often need this; see our n8n vs Make vs custom agents comparison for when the added complexity is justified.

Memory

An agent's ability to retain context — within one task, or across multiple tasks over time. Why it matters: memory is what lets an agent handle "this vendor always formats invoices this way" instead of relearning every time.

Grounding

Anchoring a model's output in verifiable external data rather than its own training-time knowledge. Why it matters: grounded answers are checkable; ungrounded ones require trusting the model's memory, which is exactly where hallucination risk lives.

Workflow vs. agent

A workflow executes a fixed sequence of steps; an agent reasons about which steps to take and in what order. Why it matters: many "AI agent" products marketed today are actually workflows with an LLM step bolted on — knowing the difference helps you evaluate vendor claims.

Autonomy levels

A spectrum from fully human-executed, through human-approved, to fully autonomous agent action. Why it matters: it's the practical framework for deciding how much oversight any given agent action needs — see our governance guide's tiered model.

Cost and infrastructure

LLM (large language model)

The underlying model — trained on large amounts of text — that gives an agent its reasoning and language capability. Why it matters: it's the engine, not the whole car; an agent is the LLM plus tools, memory, and guardrails wrapped around it.

Tokens

The units (roughly word-fragments) a language model processes and generates; usage-based pricing is typically per-token. Why it matters: tokens are the unit your inference cost is actually measured in — see our ROI model for how this shows up in a real business case.

Context window

The maximum amount of text (measured in tokens) a model can consider at once. Why it matters: it caps how much history or reference material an agent can reason over in a single step — relevant when scoping complex, document-heavy processes.

Inference cost

The per-use compute cost of running a model to generate a response. Why it matters: unlike RPA's flat per-bot licensing, inference cost scales with usage — a real line item in any agentic ROI model.

Latency

The time between a request and a model's response. Why it matters: customer-facing agentic workflows (support deflection, real-time quoting) have a latency ceiling beyond which the automation stops feeling useful.

iPaaS (integration platform as a service)

Cloud platforms (this category includes tools like n8n and Make) that connect applications and automate data flow between them without custom point-to-point integration code. Why it matters: iPaaS tools are frequently the right foundation for agentic workflows that mostly orchestrate existing SaaS apps.

RPA (robotic process automation)

Software that replays a fixed, scripted sequence of UI or API steps. Why it matters: it's the automation category agentic AI is most often compared against — see our full comparison.

Strategy

AI readiness

An organization's measurable preparedness — across data, systems, people, process, and governance — to successfully deploy AI agents. Why it matters: it's a checkable state, not a feeling; see our 25-point readiness checklist.

Generative-engine optimization (GEO/AEO)

The practice of structuring content and site infrastructure so AI answer engines (ChatGPT, Perplexity, AI Overviews) can retrieve and cite it. Why it matters: it's the emerging counterpart to SEO — see our full guide to being cited by AI.