What Actually Makes Something an "AI Agent"

The one thing that actually defines it: the loop

A single model call is a straight line: you send text in, text comes back out, done. An agent is a loop — the model takes an action, observes what actually happened as a result of that action, and decides the next step based on that observation rather than on a script someone wrote in advance.

That loop is what lets an agent do something a one-shot completion structurally cannot: adjust when reality doesn't match what it expected. A tool call comes back with an error, a search returns nothing useful, a file turns out not to exist — a single model call has no way to react to that. A loop does.

Where the real engineering decisions live

Once you have a loop, you have real design questions that a single call never raised: how much should the agent be allowed to do without checking in with a person first? What happens when a tool call fails outright — retry, try a different approach, or stop and ask? What actually stops the loop from running forever if the task never quite resolves?

None of this is exotic. It's the same territory as normal software design — error handling, permission boundaries, exit conditions — just applied to a system whose exact next move you can't fully predict in advance.

What good agent design looks like in practice

Consequential, hard-to-undo actions — sending an email, spending money, deleting something — get a human checkpoint before they happen, not a log entry after. A failed tool call is treated as information the loop reasons about, the same as any other observation, rather than a crash. And the loop has an explicit "done" condition and a step limit, so "keep going until it feels finished" is never the whole plan.