Agentic loop
The core control flow of an agent: observe → reason → act → observe, repeated until the goal is met or a stop condition fires.
Every agent in 2026 boils down to a loop. The agent looks at the current state, the LLM reasons about what to do next, a tool gets called, the result feeds back into the next iteration. Stop when the goal is met or a limit is hit.
The original ReAct (Reason+Act) paper formalized this pattern in 2022. Modern variants add planning steps, sub-agents, reflection passes, but the skeleton is the same.
What differentiates a good agent is not the loop itself — that's commodity now — but the quality of the reasoning at each step, the breadth of tools available, and the stop conditions.
Frequently asked
What is the ReAct pattern?+
ReAct (Reason+Act) is the canonical agentic loop: the LLM emits a reasoning step, then a tool call, then observes the result, then loops. Most production agents in 2026 use ReAct or a close variant.