"AI agent" is the most overloaded term in the field. It is used for everything from a chatbot with a name to software that autonomously edits your codebase, and the gap between those is enormous.
Here is what the word actually means, and how to tell a real one from a rebrand.
The actual definition
An agent is a model that can take actions and iterate, not just produce text.
Three things distinguish it:
- Tools — it can call something: search the web, read a file, query a database, send a request.
- A loop — it acts, observes the result, and decides what to do next, repeatedly, without you prompting each step.
- A goal — you give it an outcome rather than an instruction.
If it does not have all three, it is a chatbot with integrations. Which is often what you want, and is not an agent.
What the loop looks like
You ask for something. The model decides which tool to use, uses it, reads the result, decides whether it is done, and repeats.
"Find every place we reference the old pricing page and update the links." The agent searches, finds twelve files, edits each, checks nothing broke, reports back. You did not specify the twelve files. That is the difference.
Where agents genuinely work
Well-scoped mechanical work. Renaming a concept across a codebase, migrating a deprecated API, updating a value everywhere it appears. The task is unambiguous and the result is checkable.
Research with a clear question. Gathering information from several sources and synthesising it, where you can verify the sources.
Multi-step retrieval. Answering a question that requires three lookups, in an order the model works out.
The pattern: the goal is unambiguous and you can verify the result.
Where they fail
When the task needs judgement about what should be done. An agent will confidently make a large wrong change quickly, which is a worse failure than a wrong suggestion because there is more of it.
Long horizons. Errors compound. A mistake at step three propagates through steps four to twenty, and the agent has no independent way to notice.
Ambiguity. Given an underspecified goal, it picks an interpretation and commits. Rarely the one you meant.
Anything irreversible. Sending messages, moving money, deleting things, publishing. The blast radius of a mistake is the entire problem with autonomy.
The non-negotiable rules
If you use agents on anything that matters:
- Work on a branch or a copy. Never against production or your only copy.
- Review every change. The whole diff, not a summary of it.
- Have tests, or some independent check that would catch a wrong result.
- Scope tightly. "Update these twelve files" beats "improve the codebase."
- Never grant irreversible actions without a human approving each one.
The teams that get value from agents are strict about all five. The horror stories all involve skipping at least three.
How to spot a rebrand
Much of what is marketed as agentic is a chatbot with a system prompt. Useful questions:
- What tools can it actually call? A specific list, or vagueness?
- Does it loop? Can it take a second action based on the first result, or is it one shot?
- What happens when it fails? Real agents need a retry and escalation story.
- What can it do irreversibly? If the answer is "nothing", it is safer and also less agentic.
None of this means a non-agent product is worse. It means you should know what you are buying.
MCP, briefly
Model Context Protocol is an emerging standard for connecting models to tools and data — a common interface so every product does not need bespoke integration with every service.
If it succeeds, the interesting question stops being "which model" and becomes "what is it connected to." Worth knowing the term; too early to build a strategy on it.
Common questions
What is an AI agent? A model that can use tools and iterate toward a goal, rather than only responding to prompts.
How is an agent different from a chatbot? A chatbot responds. An agent acts, observes the result, and decides what to do next without being prompted each time.
Are AI agents reliable? For well-scoped, verifiable, mechanical tasks, increasingly yes. For ambiguous goals or long chains of steps, no.
Can agents replace employees? They handle bounded, checkable tasks. They do not handle judgement about what should be done, which is most of most jobs.
What is the biggest risk? Irreversible actions. An agent that can send, publish, pay or delete can make a large mistake quickly.
Do I need an agent or an assistant? An assistant, for most work. Agents earn their complexity on repetitive multi-step tasks you would otherwise do by hand.
What is MCP? A standard for connecting models to external tools and data. Early, and worth watching.