The agentic loop is the core control-flow pattern that turns a stateless LLM into an agent: the model repeatedly perceives its context, reasons about the next step, takes action (typically a tool call), observes the result, and loops until a stop condition is met. Without this loop, an LLM generates a single response and stops — with it, the system pursues open-ended goals through multi-step interaction with the world.
The four phases — Perception, Reasoning, Action, Observation (PRAO) — map to established academic framing. The ReAct architecture (Reasoning + Acting, Yao et al. 2022) first formalized this interleaving of reasoning traces with tool-use actions. In production systems the loop is also called the “Think-Act-Observe” cycle, and most agent runtimes (LangGraph, Anthropic SDK, OpenAI Agents SDK) implement it as a state machine with checkpointing between turns.
The loop’s power is also its risk: each iteration consumes tokens and wall-clock time. Without explicit termination conditions (task completion, turn budget, error threshold, stop signal) the agent can spin indefinitely, accumulating cost without progress. Production systems always pair the loop with budget guards — maximum iterations, total token caps, and wall-clock timeouts — and with escalation paths that hand off to a human operator when the loop cannot converge.
Connections
- tool-use-lifecycle — the action phase of the agentic loop depends on the tool-use lifecycle; every tool call goes through decision → execution → observation → feedback
- workflow-first-agentic-architecture — the agentic loop is the architecture of last resort; workflows (deterministic patterns) should be preferred until open-ended goals justify the loop
- system-design-patterns — the agentic loop is a system design pattern for autonomous decision-making systems