Every agent, regardless of framework, is running some version of the same three-step cycle: decide what to do, do it, and look at what happened. Frameworks give this loop different names - ReAct, a graph with a conditional edge, an "AI Agent node" - but underneath the terminology, it's the same shape every time.
Plan: deciding what to do next
In the planning step, the model looks at the goal and everything it's observed so far, and decides on the next action - which might be calling a specific tool with specific arguments, or deciding it already has enough information to answer. This is where the model's reasoning actually happens; frameworks that use the ReAct pattern (Reasoning + Acting) often have the model explicitly write out its reasoning before committing to an action, which tends to produce better decisions than jumping straight to a tool call.
Act: doing the thing
Acting means actually executing the chosen action - calling an API, running a search, querying a database, writing a file. This step is mechanical compared to planning: by the time execution happens, the decision has already been made, so this step is really just the system carrying it out and capturing whatever comes back.
Observe: looking at what happened
The result of the action - a search result, an API response, an error - gets added to what the model can see, and that updated picture is what the next planning step reasons over. This is the step that makes the loop genuinely responsive rather than just running down a checklist: a search that comes back empty changes what the model plans to do next in a way a fixed script could never account for in advance.
Plan, act, observe - then either loop back to plan again, or stop once the goal is satisfied.
The same loop, different names
If you've been through the LangGraph guides on this site, this loop is exactly what a conditional edge from an Agent node to a Tool node and back represents - the graph structure is a literal drawing of plan-act-observe. In n8n, the AI Agent node's internal tool-calling behavior runs the same cycle, just packaged as a single configurable node instead of an explicit graph. The ReAct pattern popularized in agent research papers named the "reasoning" half of planning explicitly, but it's describing the same loop.
Where the loop stops
A loop needs a stopping condition, or it's just an infinite process burning tokens - most agent implementations stop when the model itself decides the goal is satisfied and produces a final answer instead of another tool call, but production systems almost always add a hard ceiling on top of that (a maximum number of iterations) as a safety net, since a model can get stuck in unproductive loops, especially with tools that return ambiguous or unhelpful results.