“Agent” has become the most abused word in software. Everything from a search box to a macro now carries the label. This page gives you the real definition, the parts every real agent has, and the questions that show the difference in a sales meeting.
The definition
An AI agent is an AI system that has been given three things: a goal, tools, and permission to work in steps. With those, it plans an approach, acts through its tools, looks at what happened, and carries on, looping until the goal is met, the attempt fails, or a rule says a person must decide. The AI underneath is the same kind of model that powers the chatbots. The difference is what it’s allowed to do. A chatbot gives you words to use. An agent gives you finished work: the email drafted and queued, the record updated, the report put together.
The anatomy, in five parts
Every real agent, whatever the vendor calls it, has these five parts:
- The model: the reasoning engine, from any of the major providers or an open-weight model run privately.
- The instructions: a written brief setting out the job, the boundaries, the tone and the rules. In a professional deployment this is a maintained document, not a one-line prompt, and it’s where a firm’s red lines are enforced.
- The tools: the software the agent may operate, such as email, files, CRM, calendar, browser and code. Tool access sets both capability and risk, and granting it is the real deployment decision.
- The loop: plan, act, observe, adjust. This is what lets an agent recover from a failed step instead of handing you an error, and it’s where agent behaviour gets tested before anyone trusts it.
- The supervision: logs of every action, approval steps wherever the consequences are, spending caps and a defined stop. The governance guide covers this part properly, because it’s the part that keeps agents out of trouble.
The vocabulary
These are the terms you’ll hear most in agent sales pitches:
- Tool use / function calling. How a model operates software. It’s the phrase to dig into in any demo, because tool access is the real product.
- Orchestration. Co-ordinating several steps or several agents into one workflow. See multi-agent routing.
- Human-in-the-loop. A person approving defined steps. Look for it in any serious proposal, and be wary when it’s missing.
- Guardrails. The enforced limits on what an agent can do, as opposed to what it’s merely asked not to do. That difference is what governance is about.
- Autonomy levels. The ladder of trust: everything drafted for approval first, then routine actions taken with the consequential ones still gated, then acting alone with full logs. Sensible deployments climb it one process at a time rather than starting at the top.
The jargon glossary covers the rest.
What agents still cannot do
Agents have three limits that no product currently gets around:
- Carry accountability. A regulator, client or court holds your firm responsible for an agent’s work exactly as it would for an employee’s. That’s why the approval step exists, and why it can’t be outsourced.
- Escape their models’ limits. Confident mistakes are still possible, so checking belongs in the workflow, as the realistic limits guide explains.
- Define their own jobs well. Firms disappointed by agents have usually given them vague goals and broad access, the exact opposite of what works. The job description template exists to force that clarity first. If you’re being sold something “agentic”, the questions to ask in an AI demo will show how much of it is real.