Glossary · Foundations

AI Agent

An AI agent is an LLM-driven program that uses tools to complete tasks autonomously. Definition, architecture, and real SMB examples.

By Kadin Nestler · May 28, 2026 · Updated May 28, 2026
On this page

Anatomy of an AI agent

Every working agent has four parts: a model (the brain), a set of tools (the hands), a loop that lets the model call tools and read results, and a stop condition. The model is usually a frontier LLM — Claude, GPT, Gemini. Tools are anything with an API: a CRM, a calendar, a database, a web browser, a Python interpreter, another agent.

Anthropic distinguishes "workflows" (LLMs orchestrated through predefined paths) from "agents" (LLMs that dynamically direct their own processes and tool usage). Most production systems in 2026 are workflows with one or two agentic steps, not full agents — because workflows are cheaper to debug and ship.

How agents differ from chatbots

  • A chatbot returns a response to a prompt. One turn, one answer.
  • An agent can take dozens of turns against itself and external tools before returning a final answer.
  • Chatbots cost cents per conversation. Agents can cost dollars per task because they consume tokens on every loop iteration.
  • Chatbots need a person at the keyboard. Agents can run on a cron, on a webhook, or 24/7.

What SMBs actually deploy

For a 5-100 employee operator, the typical first agent is a phone receptionist (Vapi, Retell, Synthflow) that answers, qualifies, books, and writes the CRM. The second is a missed-call recovery texter. The third is a renewal or quote-tracking workflow. Each is a narrow agent with 3-7 tools, not a general-purpose assistant. McKinsey reports the agentic AI market was $5.1B in 2024 and is projected to reach $47.1B by 2030.

Risks and mitigations

  • Prompt injection: an attacker plants instructions in data the agent reads. Mitigate with input sanitization and guardrails.
  • Tool-call hallucination: the agent invokes a tool that does not exist or with bad parameters. Use strict schemas.
  • Cost runaway: hard-cap tokens per loop and iterations per task.
  • Audit trail: log every tool call. Without traces, root cause is impossible.

What it means for your business

When a vendor sells you "an AI agent," ask which tools it can call, how the loop terminates, and what the per-task cost cap is. Those three answers tell you whether you are buying a worker or a science experiment.

  • Agentic AI — Agentic AI is software that plans, acts, and uses tools to complete multi-step goals with limited human input. Definition, examples, and SMB use cases.
  • Tool Use — Tool use is when an LLM calls external APIs, databases, or code on its own. Definition, function calling, and how it powers AI agents.
  • Large Language Model (LLM) — A Large Language Model is a transformer-based neural network trained on trillions of tokens to predict the next token. Definition, key models, and business use.
  • AI Guardrails — AI guardrails are runtime rules and filters that constrain LLM behavior. Definition, types, and how SMBs should use them in production.
  • Agentic Workflow — An agentic workflow is a multi-step process driven by an AI agent that decides what to do next at each step. Definition, examples, and how to design one.
← Back to the full glossary