Interactive tool

An agent doesn't just answer — it acts, checks, and tries again.

A plain chatbot reply is one shot: you ask, it writes back from what it already knows, done. An "AI agent" is the same language model wrapped in a loop — it plans a step, calls an external tool (search the web, run code, query a database, send an email), looks at what came back, and decides whether to act again or hand you a final answer. Walk through one of those loops below, step by step.

01 — The plan → act → observe loop

Watch an agent work through a real task, one step at a time.

Pick a task, then click "Next step" to advance the loop. Each step shows what the model is thinking, the tool call it makes, the result that comes back, and what it decides to do next — until it has enough to give you a final answer.

02 — Agent vs. plain chatbot

Same model underneath, very different capability.

The difference isn't intelligence — it's whether the model is allowed to take actions in the world and loop on the results, or only generate text once from what it already learned during training.

Plain chatbot

One shot, frozen knowledge
  • ＋Generates a single reply from patterns learned during training.
  • ＋Has no live information — it only knows what was in its training data, up to a cutoff date.
  • ＋Cannot check its own answer, fetch a webpage, or run a calculation it isn't confident in.
  • ＋If it's wrong, it's wrong — there's no follow-up step to catch the mistake.
Ask it: "What's the cheapest flight to London next Friday?" It will guess a plausible-sounding price from training patterns — not today's actual fares.

AI agent

Looping, acting, checking
  • ＋Plans a step, then calls real tools — search, code execution, databases, APIs, email.
  • ＋Pulls live, current information instead of relying only on training-time knowledge.
  • ＋Observes each tool's result and can correct course, retry, or gather more before answering.
  • ＋Keeps looping until the task is actually done — not until one reply is generated.
Ask it the same thing: it calls search_flights(...), reads back real fares, compares them, and only then answers — or books the ticket if you told it to.
03 — Check your understanding

Five questions on agents and tool use.

Pick an answer for each question to see if you're right.

Score: 0 / 5
04 — Key takeaways

What actually makes something an "agent."

None of this requires a smarter model — it's an architecture choice: give the model tools, a loop, and the judgment to decide when it's done.

Related tools

Keep exploring how AI systems are actually put together.

Interactive tool

The AI Value Chain →

Tool use is one link in a longer chain — see everything that sits between a raw model and a working product.
Interactive tool · coming soon

Retrieval vs. Memory →

An agent's "observe" step often pulls from a retrieval system — learn how that's different from a model simply remembering.
Interactive tool

AI Use-Case Triage →

Once you know what an agent can do, use this board to sort which of your own use cases actually need one.
05 — Go deeper

Deciding when a loop is worth the risk.

Agents are powerful precisely because they can take real actions — which also means a bad plan or a misread tool result can compound across iterations instead of stopping at one wrong reply. Knowing where to put guardrails, approval steps, and budgets on the loop is exactly the kind of design work we help teams get right before they ship an agent.

Talk to us about your AI roadmap