ReAct

You might say

Can the AI check the result after each step before deciding what to do next?

A task pattern where an agent decides, acts with a tool, observes the result, and decides againWhen a login button does nothing, an agent can inspect the button first. If the click handler already exists, that observation changes the next action to inspecting the network request and its 401 response.

Know first

Tool Calling

Reason + Act

Why inspect the network request when the button file is correct?

tool call

1read_file("src/login-button.jsx")

Initial hypothesis: the button may lack a click handler, so read the actual file first.Initial hypothesis

How a real result changes the next action

Verify one concrete assumption: For a button that does nothing, first check whether it has a click handler.

Read the actual result: The file shows that the click handler already exists, disproving the first assumption.

Change the investigation: A 401 after the click moves the next check to authentication. This is task state, not private chain of thought.

Quick checkChoose the best answer

An agent assumes a login button has no click handler, but the file shows that the handler already exists. What should happen next in a ReAct pattern?

Drop the original assumption and inspect the network request triggered by the clickRewrite the click handler anyway because the original plan blamed the buttonDeclare the button correct and stop without checking what happens after the click

You can say this to an AI Agent

Use a ReAct pattern to diagnose this issue. State one thing to verify, perform one read or check, update the task-level conclusion from the result, and then choose the next action. If evidence disproves the original assumption, change direction instead of continuing the same edit.

Learn next

1Agent Loop→2Tool Calling→3Stateless Request→

Want GPT or Codex?VibeHubGPT support

GPT-6 Astra

Build in 3D. Put your computer to work.