The agentic loop is the core control cycle of an AI agent: observe the current state, reason about the goal, act (call a tool, write output), observe the result
It is an AI agent's core control cycle: observe the current state, reason about the goal, act (call a tool or write output), observe the result, and repeat until the task completes or an exit condition triggers. Loop design largely determines agent quality.
Runaway loops, where an agent burns tokens repeating failed strategies, are the canonical failure mode. They are bounded with step limits, budget checks, progress checks, reflection steps that ask whether the last action worked, and escalation paths when the agent is stuck.