“Agent” has become the most abused word in software. Everything from a search box to a macro now carries it. This page gives you the real definition, the anatomy every genuine agent shares, and the questions that expose 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. Given those, it plans an approach, acts through its tools, observes what happened, and continues, looping until the goal is met, the attempt fails, or a rule says a human must decide. The underlying intelligence is the same class of model that powers the chatbots; the difference is posture. A chatbot’s output is words for you to use. An agent’s output is work: the email drafted and queued, the record updated, the report assembled.
The anatomy, in five parts
Every genuine agent, whatever the vendor calls it, is 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 defining 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 is where a firm’s red lines get enforced.
- The tools: the software the agent may operate. Email, files, CRM, calendar, browser, code. Tool access defines both capability and risk; 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 is where agent behaviour is tested before anything is trusted.
- The supervision: logs of every action, approval gates where consequences live, spending caps, and a defined stop. The governance guide treats this part properly, because it is the part that keeps agents out of trouble.
The vocabulary
- Tool use / function calling. The mechanism by which a model operates software; the phrase to probe in any demo, because tool access is the real product.
- Orchestration. Coordinating multiple steps or multiple agents into one workflow; see multi-agent routing.
- Human-in-the-loop. A person approving defined steps: the phrase to look for in any serious proposal, and to distrust when it is missing.
- Guardrails. The enforced limits on what an agent may do, as opposed to what it is merely asked not to do; the difference is the whole subject of governance.
- Autonomy levels. The ladder of trust: everything drafted for approval, then routine actions taken with consequences still gated, then acting alone with full logs. Sane deployments climb it per process rather than starting at the top.
The jargon glossary covers the rest of the dialect.
What agents still cannot do
- Carry accountability. A regulator, client or court holds your firm responsible for an agent’s work exactly as for an employee’s, which is why the approval gate exists and cannot be outsourced.
- Escape their models’ limits. Confident error remains possible, so verification belongs in the workflow, per the realistic limits guide.
- Define their own jobs well. The firms disappointed by agents almost always gave them vague goals and broad access, the exact inverse of what works; the job description template exists to force the clarity first.
Where I fit in
I build and run agents daily, under written specifications, logged actions and human gates; the Software Factory page shows the system and the Labs page shows its output. If you are being sold something “agentic” and want the claims translated into what it will and will not actually do in your firm, that is a conversation I have often, and the Automation Audit makes it systematic.