Definition
Designing the feedback cycles around an agent so action, verification, operational triggers, and system improvement have explicit state, evidence, limits, and stopping conditions. LangChain's proposed taxonomy nests four levels: a core agent loop does work, a verification loop checks an attempted result, an event-driven workflow connects the agent to an operating system, and an outer improvement loop changes the harness using evidence from prior runs.
The useful insight is not the number four. It is that each loop has a different job and therefore needs a different completion rule, authority boundary, cost budget, and evidence standard. The outer loop should improve the inner loops without sharing their authority or changing their evidence.
Why it is contested
The phrase entered current agent-engineering discussion through LangChain's 2026 writing, but it is not a settled discipline or a claim that feedback control is new. The proposed four-level taxonomy is one useful framing, not a universal decomposition. Other systems draw the boundaries differently or use established control, workflow, and reliability language for the same mechanisms. This glossary retains the term when it helps separate doing work, checking work, responding to events, and changing the system that does the work.
Distinguish it from nearby terms
An agent loop is the local model-and-tool cycle. A verification loop evaluates an attempted result. An improvement loop changes the harness using evidence from prior runs. Graph engineering makes nodes, state, and transitions explicit; a loop is a cyclic graph, so the terms describe overlapping views rather than competing architectures.
Check your understanding
A failed test sends code back to the same agent without a retry budget, while a self-improvement process can rewrite that test. Both loops are present, but their authority is entangled and neither provides independent evidence.