Agents, part 1: what they are and where they fit

"Agent" is everywhere in AI marketing — and almost nowhere in plain English. A chatbot is not an agent. An autocomplete button is not an agent. An agent is closer to a junior coworker with a keyboard: it reads a goal, thinks in language, uses software on your behalf, checks what happened, and tries again. Part 1 is the map; part 2 is the wiring.

Start with the stack, bottom to top

Think of four layers, not one magic box. Machine learning is the old foundation: systems that learn patterns from data — spam filters, fraud scores, image tags, demand forecasts. They predict or classify; they do not usually hold a conversation. Language models are a newer layer built on that same machinery, trained on huge amounts of text (and often code) so they can read and write fluently. A prompt is what you send the model: a question, a draft to fix, a set of rules, or a task description. The model answers once per turn unless you build a loop around it. An agent is that loop: model plus goal plus tools plus memory (sometimes) plus permission to act, running until the job is done or someone stops it.

What a prompt does vs what an agent does

A prompt is a single ask: "Summarize this PDF," "Write three subject lines," "Explain GRI rule 3(b)." You get text back. You copy, paste, click, or decide. An agent gets a mission: "Book the cheapest refundable flight that lands before 5pm," "Open a PR that fixes the failing test," "Reconcile these 400 SKUs against the HTS schedule and flag anything below 90% confidence." The system breaks that into steps, calls tools (browser, calendar, shell, API, spreadsheet), reads the results, and keeps going. You are not clicking Send after every micro-move — you are supervising outcomes.

Why the word feels slippery

Vendors slap "agent" on anything that feels smart. Copilots that only suggest text in your editor are assistants, not agents. Workflow automations with fixed if/then rules are not agents either — there is no model reasoning through novel steps. Real agents sit in the middle: flexible like a chatbot, but wired to the outside world like automation. That flexibility is the point — and the risk. When agents get credentials, the failure mode stops being a bad paragraph and starts being a wrong click at scale.

A simple picture

You → goal → agent → model (thinks in language) → tools (does things) → world → results back → agent adjusts → repeat. Machine learning still lives underneath (ranking, retrieval, safety classifiers). Language models are the reasoning surface. Prompts steer one turn. Agents wrap many turns into one job. That is the whole cheat sheet for part 1.

What to watch in the directory

Browse Agentic AI for products where the loop is the product — not a chat window with a marketing rename. If you want the generative layer underneath, start with our Generative AI explainer. When agents go wrong in the wild, our Hugging Face incident report is a concrete example of tool access plus too much autonomy.

Continue with part 2

Ready for the wiring? Part 2 covers tools (APIs, MCP, browsers), memory, planning, multi-agent handoffs, guardrails, and where human approval should sit in the loop.

The point

An AI agent is a language model that has been given a goal, access to tools, and permission to keep trying until the task is done. Most confusion comes from using "agent" for all three of those layers at once. When a product calls itself an agent, ask which model it runs on, what it is trying to accomplish, and which tools it is allowed to use. Those answers tell you more than the label.