Agent Loop

Process agents published

Also known as: Reasoning Loop, Action Loop, Observe-Think-Act Loop

Definition

The iterative cycle an agent follows: perceive the current state, reason about what to do, execute an action, observe the result, and repeat until the goal is achieved or a termination condition is met. The agent loop is what distinguishes agents from single-turn inference—it enables multi-step problem solving where each step's output informs the next step's input.

What this is NOT

  • Not a single inference call (loops require multiple iterations)
  • Not a batch process (loops are reactive to intermediate results)
  • Not recursion in the programming sense (though implementation may be recursive)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

The core execution pattern of LLM agents: prompt the model, parse the response for an action, execute the action (often a tool call), append the result to context, and loop. Common implementations include ReAct, function-calling loops, and plan-then-execute patterns.

Sources: ReAct paper (Yao et al., 2022), LangChain AgentExecutor implementation, Anthropic tool-use documentation

control-theory

A feedback loop where a controller observes system state, computes control signals, applies them to the system, and measures the new state. The classic observe-orient-decide-act (OODA) loop.

Sources: Control systems engineering literature

Examples

  • ReAct loop: Thought → Action → Observation → Thought → ...
  • Function calling loop: prompt → function_call response → execute function → append result → prompt
  • Self-correction loop: generate → evaluate → regenerate if quality is low
  • Research loop: query → retrieve → assess sufficiency → query again or synthesize

Counterexamples

Things that might seem like Agent Loop but are not:

  • A single GPT-4 API call that returns a complete answer
  • A RAG system that retrieves once and generates once
  • Batch inference over a dataset

Relations

  • requires agent (Agents execute through agent loops)
  • requires reasoning (The loop includes a reasoning step)
  • overlapsWith tool-use (Tool execution is often the action in the loop)
  • overlapsWith planning (Some loops include explicit planning steps)

See Also

Implementations

Tools and frameworks that implement this concept: