What is an "AI agent"?

1. Assistant or agent: the real difference

When you chat with Claude, it's an assistant: you ask, it answers, you take back the wheel. An agent is one step up: you give it a goal, and it chains the steps on its own — it searches, decides, acts, checks its result, and tries again if needed, until it delivers.

The key ingredient is autonomous looping + access to tools to take action.

2. What it's actually useful for

In my setup: a news brief that composes and publishes itself every morning without me touching it. And agents in progress for sorting emails or watching numbers. Honestly, some are running, others I've paused — because exactly…

3. The essential guardrail

…an agent without constraints is a robot let loose. I had one that kept restarting in a loop and burning through its budget every day without producing anything usable. The lesson, which became a rule for me: every agent needs four things before being launched —

1. an explicit goal ("we'll know it's done when…"),
2. a hard budget cap,
3. a limited frequency (not a permanent loop),
4. a measurable success criterion.

4. Where to start

Small. One agent, one task, and prove it's reliable before launching a second. The "agents everywhere" big bang is the best way to break everything.