What Makes an AI Agent

“Agent” has become one of those words that means whatever the slide deck needs it to mean. The definition I keep coming back to is simple: an AI agent doesn’t just reply. It perceives something, reasons about a goal, plans, and then acts across several steps and tools until the goal is met (or it gives up sensibly).

The four stages

Picture a travel concierge running on your phone. Your evening flight is cancelled.

Stage	What the concierge does
Perception	Notices the cancellation in your booking feed
Reasoning	Works out that the goal hasn’t changed: you still need to arrive tonight
Planning	Compares alternative flights and trains against time and cost
Action (multi-step)	Books the best combination and messages the hotel about a late check-in

A chatbot would have told you your flight was cancelled. The agent fixed it.

It’s a loop, not a prompt

Strip away the LLM and this is the same loop as Reinforcement Learning: an agent observes an environment, takes an action, sees the result and adjusts. The difference is that a language model does the reasoning and tools do the acting — APIs, a browser, a shell, a code interpreter. Eureka is a lovely worked example: an LLM writes code, runs an experiment, reads the results and rewrites the code.

The pattern most production agents converge on looks like this:

1. Propose — decide the next step from the goal and what’s known so far.
2. Act — call a tool.
3. Observe — read the tool’s output, errors included.
4. Reflect — was that progress? Update the plan.
5. Stop — on success, on a budget, or when a human needs to decide.

The trade-offs

• Autonomy vs control. Every extra step an agent takes on its own is a step nobody reviewed. Good agents have explicit budgets, stopping criteria and points where they hand back to a person.
• Tools are the real interface. An agent is only as good as the contracts it can call. Vague APIs, missing error messages and undocumented side effects hurt agents even more than they hurt people, because agents won’t ask in Slack when the docs are wrong.
• Observability isn’t optional. You need to see what the agent perceived and why it chose an action — the same thinking as Observability for any distributed system.

In practice

I run a small agentic loop every evening: an assistant reads my day’s raw notes, triages what’s worth keeping, plans where each piece belongs and then edits many notes in one go. It’s the four stages above, pointed at a knowledge base instead of a travel booking — and it’s the reason this garden gets fed at all. The lesson from building it matches what platform teams are discovering: the agent itself is the easy part; clear rules, safe defaults and a log of what it did are what make it trustworthy. That’s why I think of agents as a new kind of platform user (see Where Platforms Fit: DDD Subdomains).