AI Agents, Explained

An AI agent is a system built on top of a language model that can take a goal, break it into steps, use tools to act on the world, check its own results, and keep going without a human approving every move. That's the practical difference from a chatbot: a chatbot answers one message at a time and waits for you; an agent keeps working across many steps toward an outcome.

Agent vs. chatbot: what actually changes

A chatbot's job ends when it returns text. An agent's job ends when the task is done. Three things usually separate the two:

• Tool use — the model can call a search engine, run code, read a file, hit an API, or click through a web page, not just generate text.
• Planning across steps — the system decides what to do next based on what happened after the previous step, instead of a single input-to-output pass.
• Self-checking — many agents review their own output (did the tests pass? did the API call return an error?) and retry or adjust before finishing.

None of this requires a fundamentally different model — it's mostly scaffolding: a loop, a set of tools, and rules for when to stop.

What people are actually using agents for right now

• Coding agents that open a repo, make a change across multiple files, run the test suite, and fix what breaks — this is the single most mature agent use case today, because "did the tests pass" is an easy, unambiguous signal for the agent to check itself against.
• Research and browsing agents that read multiple sources, cross-check facts, and return a synthesized answer with citations instead of one search result.
• Workflow automation agents — often built on tools like n8n or similar orchestration platforms — that watch for a trigger (a new email, a form submission, a webhook) and carry out a multi-step business process without a person in the loop for the routine cases.
• Customer support agents that can look up an order, check a policy, and issue a resolution — escalating to a human only when a rule says to.

What tends to go wrong

Agents fail differently than chatbots do. A chatbot gives one wrong answer; an agent can act on a wrong assumption for several steps before anyone notices, which is why the systems that work well in production usually have:

• Bounded scope — a narrow, well-defined job rather than an open-ended one.
• Checkpoints — points where the agent must produce a verifiable result (tests passing, a schema-valid output) before continuing.
• A human-in-the-loop for anything costly or irreversible — sending money, deleting data, or emailing a customer usually still gets a review step, even in otherwise autonomous systems.

What it costs to run one

Agent workloads are usually priced the same way as any other LLM usage — per input and output token — but they use a lot more tokens per task than a single chat message, because every tool result, every intermediate step, and often the full conversation history gets fed back into the model on each turn. A single agent "run" that completes a multi-step task can easily use 10-50x the tokens of one chatbot exchange. That makes the per-token price of the model you pick a bigger lever than it is for chat use cases — see the live pricing comparison for current rates across providers, and the AI Model Picker if you want a recommendation based on your own budget and context-window needs.