Agent Loop

Pain Signals — You Need This Pattern When:

• A single LLM call cannot solve the task — it requires multiple steps with intermediate observations
• The number of steps is not known in advance — the agent must decide when it is done
• Each step depends on the result of the previous step (not a fixed pipeline)
• The task requires exploration — reading files, searching, querying databases, trying approaches
• You need an LLM to plan and execute a multi-step workflow autonomously

This is the most complex LLM pattern. Use it only when Prompt Chaining (fixed pipeline) and Tool Use (single-round) are insufficient.

Core Principle

An agent is a state machine that loops through observe-think-act cycles until a termination condition is met. The LLM provides the "think" step; your system provides the "observe" and "act" infrastructure.