What an AI agent actually is
An agent is a loop. It looks at where things stand, decides what to do next, does it using a tool, reads the result, and goes again until the job is done or it needs you.
Three things make it an agent rather than a clever chatbot: it acts instead of only answering, it uses tools to reach your files, your inbox and your systems, and it works from a goal rather than a single instruction. We unpack that properly in AI Agents, actually explained.
Two words you will meet immediately and should not confuse. A tool is one capability, like “search the CRM” or “send an email”. A skill is a written procedure the agent picks up when the job calls for it, closer to an SOP than to software. Neither is an agent. An agent is the loop that uses them.
Getting this straight early matters more than it sounds. We designed our own LinkedIn build as a multi-agent system and only discovered, when we applied our own test to it, that it was one model changing hats inside a single conversation. Sophisticated, useful, and not what we had been calling it. The fix took an afternoon once we knew.
The four big vendors do not agree what an agent is
This explains most of the confusion in this market, and almost nobody says it out loud. Anthropic, Google, OpenAI and Microsoft all publish a definition of an AI agent. All four are defensible. None of them match.
| Vendor | What they say an agent is | What that actually tests |
|---|
| Anthropic | The model directs its own process and tool use, within the limits you give it. Explicitly contrasted with a workflow, where the path is decided in advance by whoever built it | Control flow. Who decides the next step |
| Google | An application that pursues a goal by observing the world and acting on it using tools, built from a model, tools and an orchestration loop | Anatomy. Does it have the parts, including the loop |
| OpenAI | Systems that independently accomplish tasks on your behalf | Independence. Who does the work |
| Microsoft | Expert systems acting on behalf of a process or a company, in three tiers: retrieval, task and autonomous | Job to be done. The bottom tier is a chatbot |
The most useful fact in that table: a Microsoft “retrieval agent” is what Anthropic would call an AI feature. Same thing. One vendor calls it an agent, the other refuses to. So somebody who built in Copilot Studio has been correctly told by their vendor that they own an agent, and will be told elsewhere that they do not.
That is not anyone lying. It is two published definitions disagreeing, and it needs more care than a correction.
Where all four agree: it has to act on the world through tools. A model that only produces text is out, unanimously. That is the floor worth asserting when the definitions collide.
We use Anthropic’s, because it is the only one you can actually check rather than argue about. It is also the strictest, which is why we tell people they do not need an agent more often than most.
Build the crude version first
The build loop that works, in order:
- One task, end to end, badly. Get something running the whole way through before you improve any part of it. A crude agent that finishes teaches you more than a beautiful one that does half the job.
- Watch it work. Run it on real examples and read every output. You are not testing quality yet, you are finding out what you forgot to tell it.
- Widen slowly. One capability at a time. Resist giving it more tools: an agent with four sharp tools beats one with fifteen it has to choose between.
- Add the guardrails last, once you know what actually breaks rather than what you imagined might.
A note on cost, because it surprises people. Credit-priced builders charge for every attempt, including the ones that fail while you are still learning the tool. Budget for the learning, not just the running.
What breaks first
Roughly in the order you will meet it:
- It invents things. A file that does not exist, a person who does not work there. Usually means your instructions were looser than you thought.
- It loops. Runs forever, or stops far too early. It needs a clearer definition of done.
- The tone is wrong. Solvable, but this is the one that needs a human reading the output for a while.
- It reaches something it should not. Scope its access rather than trusting it to be careful.
The failure that actually costs you is none of these, because they are all loud. It is the silent one: the agent that keeps running and quietly produces nothing useful. Four of our own pipelines did exactly that for weeks. Nothing alerted, everything looked healthy, and we found out because someone complained.
That lesson is the whole of the People page, and it is the one most people skip.
