What is an AI agent?
An AI agent is a software system in which a language model decides, step by step, which actions to take — calling tools, reading the results and continuing — until a goal is met or a rule stops it. The model chooses the next action. A chatbot chooses the next sentence. Everything else about agents — cost, risk, governance — follows from that one difference.
The definition, unpacked
A language model is the decision-maker. It reads text (the request, the history, tool results) and writes text (a reply or a structured tool call). It has no memory of its own between calls, no access to your systems, and no ability to act — those are provided around it.
Tools are functions your code exposes to the model with a name, a description and a schema: get_availability, create_booking, send_payment_request, handoff_to_human. The model can ask for one to be run; it cannot run it. Your orchestrator runs it — after checking guardrails.
The loop is what makes it an agent. Request → model → tool call → result → model → … → reply. In the booking demo on this site, one customer message produces eight model calls and six tool calls before the conversation ends with a confirmed appointment and a human handoff. A chatbot would have produced one reply per message and no bookings.
The stop condition matters as much as the loop: goal met, customer asks for a human, confidence low, budget of steps exhausted, or a guardrail fires. An agent without a stop condition is a bill without a ceiling.
What is not an agent
- A decision-tree chatbot (menus, "press 1 for balance", intent buttons). Predictable and cheap; it cannot handle "I need a check-up this week, afternoons".
- A chat interface over a model with no tools. Fluent, sometimes helpful, but it cannot verify a price, hold a slot or send a payment prompt — it can only say that it did.
- A workflow automation (Zapier, n8n, native CRM rules). Trigger → condition → action on structured data. Deterministic, auditable, the right tool for half of what people ask agents to do. See agent vs automation.
- RPA bots that replay clicks. Brittle, but deterministic; they do not decide anything.
The four-question test
Before you scope an agent, answer these in order. The first "no" tells you what to build instead.
- Do the inputs vary in ways a menu cannot capture? Free text, voice notes, photos of a form, mixed English-Shona. If not → a chatbot or a form.
- Does completing the job require an action in a system — a booking, a quote, a ticket, a payment request? If not → a chat interface with retrieval is enough.
- Can every action be made reversible or confirmable? Idempotent writes, a "reply YES to confirm" step, a human approval for anything financial. If not → the agent may only recommend; a person acts.
- Can you write down the policy the agent must follow — scope, refusals, escalation rules — and test it against a golden set? If not → you are not ready; write the policy first.
Four yeses: build an agent, and price it on the costs page. A no at question 1 or 2: read agent vs chatbot. A no at 3 or 4: read risks and governance first.
Where each part sits
The orchestrator is ordinary software — a small service that holds the conversation, calls the model, validates tool calls against a permission list, executes them, and records everything to a trace. The model is stateless; memory is a database the orchestrator reads and writes. Guardrails are code in the orchestrator, not sentences in the prompt. The full architecture, with the parts you must build versus buy, is on how agents work.
Frequently asked
Is ChatGPT an AI agent?
Not by itself. A chat interface with no tools chooses sentences, not actions. It becomes an agent when it is given tools it can call and a loop that executes them — which some products now bundle in.
Does an AI agent need the internet?
Yes, in practice: the model runs on a provider's servers, and channels like WhatsApp are cloud services. A self-hosted model on local hardware is possible but expensive and less capable; it still needs connectivity to reach WhatsApp or a phone network.
Can an AI agent take payments in Zimbabwe?
It can request them: through Paynow it can send an EcoCash or OneMoney payment prompt to a phone and poll for the result. It should not hold card details or move money without a confirmation step and an audit log.
Can an agent speak Shona or Ndebele?
Current large models handle Shona and Ndebele text to a usable but imperfect degree; speech recognition and synthesis in both languages vary widely by provider and must be tested with your own callers before launch.
Is an AI agent the same as automation?
No. Automation runs fixed rules on structured inputs and always does the same thing. An agent interprets unstructured input and chooses actions; it is wrapped in automation-style guardrails so its choices stay inside policy.