What is an AI agent?

The one question that separates an agent from a chatbot

Who decides what happens next?

In a chatbot, nobody does. You send a message, the model produces one reply, and the turn is over. If the reply needs a follow-up, you write it. The model never acts on the world; it only writes text.

In an agent, the model decides. Given a goal, it picks an action, such as searching the web, running a piece of code, opening a page or editing a file, gets the result back as text, reads it, and picks the next action. It keeps going until it judges the goal met or a step limit stops it. The model is the loop's controller.

Anthropic's engineering guide "Building effective agents", published in December 2024, draws the line in the same place: workflows are systems where the model runs inside code that fixes the path, and agents are systems where the model directs its own process and tool use. That guide is also blunt about which of the two to build first, and its answer is the workflow.

Chatbot, workflow, agent: the same model, three different loops

All three can run on the same language model. What differs is what surrounds it.

The workflow column is where most useful automation lives. A pipeline that summarises a document, then extracts the dates, then drafts a reply is three model calls in a fixed order. It is reliable because the developer decided the path. It isn't an agent, whatever the landing page says, and calling it one obscures the real trade-off: a fixed path is predictable and cheap, a model-chosen path is flexible and neither.

What the loop looks like, step by step

A concrete run. The goal: "Find out whether our competitor changed their pricing this month and write me a two-paragraph summary."

1. The model reads the goal and decides the first action: a web search for the competitor's pricing page.
2. The search tool returns a list of results as text. The model reads it and picks one to open.
3. The browser tool returns the page. The model extracts the current prices.
4. The model decides it needs the old prices, searches for an archived copy, and opens it.
5. It compares the two, judges that it has enough, and writes the summary.

Five steps, four tool calls, six model calls. Every step is a full answer from the model, with the whole history so far as input. That last detail is the cost driver: by step five the model is re-reading four pages of search results and page text on every call. What are tokens in AI explains why input that grows every turn is the expensive shape.

And every step is a place to go wrong. If the search in step 1 returns the wrong company, steps 2 to 5 confidently summarise the wrong pricing. The loop compounds errors the same way it compounds progress. A step cap and a human checkpoint before anything is sent aren't optional.

What an agent run costs against a single answer

The arithmetic is simple and people skip it. One standard answer, 1,000 tokens in and 500 out, has a list price per model in the AI model cost index, with OpenRouter rates fetched on 2026-08-20. An agent run of 20 steps is at least 20 such answers, and in practice more, because the input grows each step.

Twenty steps on Claude Sonnet 5 is $0.14 at the floor, against $0.007 for one reply; on Claude Opus 5 it is $0.35, and run every hour that is about $250 a month before the growing-input multiplier. Whizi credits read the same arithmetic off the Auto ladder above. There, 100 credits is 100 quick questions on the house model or 5 reasoning turns on GPT-5.6 Sol. So a 20-step loop at the reasoning rung is 400 credits, about a sixth of a 2,500-credit Pro month. On the house model it's 20. That's why the right design routes the planning steps to a strong model and the routine steps to a cheap one. It's also why a multi-model workspace, not a single-model subscription, is the natural home for agent work.

Which products are actually agents

Coding assistants that edit files and run tests are agents in the strict sense: given a bug report, they read code, make a change, run the test suite, read the failure, and try again. The loop is real and the model steers it.

Deep research modes in ChatGPT, Gemini and Perplexity are agents: they search, read, search again based on what they found, and stop when they judge the question covered. The step count varies per question, which is the tell.

Most "agents" in product menus are personas or workflows. A row in a model picker labelled as a tax agent or a study mentor is usually a fixed system prompt plus a tool or two. Whizi's own Agents section is honest about this: the Canadian CPA row is a preset that forces a web search on every turn, capped at seven searches, with a larger token budget of 55,000 in and 40,000 out. That is a persona with a tool attached and a bigger budget, and it is useful precisely because the path is fixed. It isn't a loop the model steers, and the Auto routing guide explains what actually picks the model behind each message.

The vocabulary is drifting toward calling everything an agent. The question to ask a product is the one at the top of this page: who decides the next step? If the answer is "a script" or "nobody", you are buying a workflow or a chatbot, and that may be exactly what you want.

When an agent is the right tool, and when it is not

An agent earns its cost when the path cannot be known in advance: research where the second search depends on the first result, debugging where the fix depends on the test output, browsing where the next click depends on the page. The flexibility is the product.

An agent is the wrong tool when the path is known. Summarise, extract, translate, reformat, classify: these are one call or a fixed sequence of calls, and wrapping them in a loop adds cost, latency and a new way to fail without adding capability. The Anthropic guide's advice stands: start with a single call, move to a workflow when one call is not enough, and reach for an agent only when the workflow keeps needing a branch you cannot predict.

If you can draw the steps on a whiteboard before you start, don't build an agent.