"Agent" gets used loosely enough now that it's worth having a quick way to check if something actually is one, or if it's just a chatbot or a script wearing the label. Here's a four-question test that usually settles it.
1. Does it take action, or just suggest one?
A tool that tells you what to do isn't the same as a tool that does it. If the system writes a draft and waits for you to send it, that's an assistant. If it sends the email itself as part of finishing the task, that's a real action — the kind of thing the actual line between an assistant and an agent comes down to.
2. Does it decide what to do next, or wait for you to tell it?
This is the core of it. A script runs the same steps every time, in the same order, no matter what it finds along the way. An agent looks at what just happened and picks its next step based on that — check the account, then decide whether to escalate or resolve, based on what it found.
If every step is predetermined regardless of the input, you're closer to a fixed workflow — which might genuinely be the better fit, depending on how much the task actually varies.
3. Does it use tools on its own, or only when you trigger them?
An agent typically has access to a set of tools — search, a database, an API, a calendar — and chooses which one to call based on the task, without you specifying which one to use each time. If a human is picking the tool and the "agent" just executes that one action, the decision-making is happening outside the system, not inside it.
4. Does it remember and use information from earlier steps to change its next move?
This one trips people up because "remembering" the current conversation isn't the same as having memory that changes future behavior. If a system used something it learned three steps ago to make a different decision now, that's real state being carried forward. If it's just re-reading the same conversation transcript every time without acting on anything from it, that's not the same thing. The distinction between short-term context and actual memory is where a lot of "agents" quietly fall short.
Scoring it
Four yeses, and you're building or using a genuine agent — something that plans, decides, and acts with real autonomy over the process. Mostly no's usually means assistant, chatbot, or a fixed automation, and none of those are worse. They're just a different, often simpler and more predictable tool for a task that doesn't need judgment calls.
The test isn't really about the label. It's a fast way to check whether you're about to build something more complex than the task actually requires — or, just as often, less capable than the task actually needs.
More on how these systems actually differ over at Procwire.
Top comments (0)