Skip to main content

COMPEL Glossary / agent-loop

Agent loop

The iterative pattern in which a model plans, calls tools, observes the results, and replans — continuing until a goal is met, an iteration budget is exhausted, or a human intervenes.

What this means in practice

The agent loop is the defining architectural element of agentic AI; its boundaries (tool set, iteration limit, halt conditions) are the primary levers for governance.

Synonyms

agent execution loop , agentic loop

See also

  • ReAct — An agentic prompt pattern interleaving reasoning steps (Thought) and action steps (Act — typically a tool call) over multiple turns, with observations from each action feeding the next reasoning step.
  • Tool use / function calling — A prompt pattern authorizing the model to request named functions with structured arguments — searching the web, reading a database, calling a calculator, triggering an API — rather than generating all answers from its weights.
  • Excessive agency — A failure mode in which an LLM has been wired into tools and permissions whose blast radius exceeds what its supervision and validation logic can safely bound.
  • Agent runtime — The execution substrate that hosts the agent loop, its tool calls, state, and recovery logic — e.g., LangGraph, CrewAI, AutoGen, or OpenAI Agents SDK.

Related articles in the Body of Knowledge