An AI agent is a system that takes a goal and produces a finished outcome, by planning steps, calling tools to take real actions, observing the results, and adjusting until the goal is met. A chatbot answers; an agent finishes the work.
Tools: how agents act on the world
On their own, models only produce text. Tools give an agent capabilities: look up an order, send an email, query a database, browse a page. A well-designed tool has:
- A clear name and description the model reads to decide when to use it.
- A typed input schema that's validated before your code runs.
- A handler you control, returning structured output.
The agent loop
A typical run looks like: plan → act (call a tool) → observe → repeat → verify. The best platforms let you watch this loop, streaming the plan, each tool call, and each result, so agents are transparent and debuggable instead of a black box.
Designing agents you can trust
- Keep humans in the loop for consequential or irreversible actions.
- Scope tools tightly, never expose more capability than a task needs.
- Verify before shipping, check quality, policy and rights, not just that text was produced.
See Agents & tools for the full pattern, including streaming events and verification.