Agent Loop

You might say

The AI stops halfway through. Can it keep working on its own and only ask me when it gets stuck?

The runtime mechanism that repeatedly calls a model and tools, then decides whether to continue, pause, or stopWhile fixing a page button, a product feeds each read, edit, and check result back to the model and stops after a passing check, a limit, or a request for confirmation. It does not treat “done” in model text as proof.

Know first

ReAct Stateless Request

Agentic Loop

How a task is controlled through continue, pause, and stop

Agent runtimeRound 1 · illustrative limit: 4 roundsContinue

Tool permission: project filesHigh-risk actions require confirmation

Model proposes reading a file

read_file("nav.jsx")

The model proposes the next step; the Agent runtime executes it, saves results, and decides whether to continue, pause, or stop.

How a runtime controls the work

Track progress：The runtime keeps the goal, completed steps, tool results, and remaining attempts.

Allow only approved actions：It validates each response or tool request before running the allowed action and recording its result.

Stop on conditions：A passing check, error, approval request, or maximum number of turns pauses or ends the task.

Learn next

1ReAct→2Tool Calling→3Rate Limit→4Token Usage & Cost→