AI Agent

You might say

I don't want to direct every step. Can I give it a goal and let it figure out how to finish it?

An AI system that can plan steps, use permitted tools, and act on results toward a goalTo fix a page, an agent can read the relevant file, edit it, run a check, and report what passed. It only has the tools, permissions, and context the product gives it.

Know first

AI Application Basics Tool Calling

Agent

How an agent moves from a request to checked work

AI Agent

You

Fix the navigation overflow on small screens, then verify it yourself.

Agent runtime

Permission: project filesHigh-risk actions need confirmationStart

Readread_file("nav.jsx")Round 1 round
InspectObservation: 84 lines; no small-screen media queryRound 1 round
Changepatch nav.jsx: horizontal scrolling on small screensRound 2 round
Runnpm run build → ✓ passedRound 2 round

Task recorded: fix small-screen navigation overflow and verify it when finished.

A goal can span several steps. The Agent records the task and success criteria before rushing to an answer.

What matters is not that an Agent can chat, but that it can act from results within controlled boundaries and knows when to stop.

What makes an agent more than a chat reply

Set a goal and constraints：The product gives the agent a task, protected scope, and the information it needs.

Use permitted tools：The model can request reads, edits, or checks, but the runtime controls what it may do.

Report from evidence：Real tool results, file changes, and tests determine whether the task is actually complete.

Quick checkChoose the best answer

Which example is best described as an AI agent rather than a single response or a predefined workflow?

It reads the error and relevant files, then uses test results to decide whether to investigate, edit, or ask for human judgmentA model receives one article and returns one summary without external information or follow-up actionCode always runs classify, translate, and send in that order, and the model cannot change the path or choose another tool

You can say this to an AI Agent

The cause of this sign-in failure is unclear. Read the error and relevant code first, then choose the checks you need. Run the relevant test after each change. Do not alter unrelated features or read real secrets. Stop when tests pass and show the evidence; ask me before making a product decision.

Learn next

1Tool Calling→2Agent Loop→3Sub-agent→